Bango - Help & Guides

route

The Big Picture

Bango follows the standard systematic review process: you import search results from academic databases, remove duplicates, define what you are looking for, and then let AI help you screen each article's title and abstract. You always have the final say on every decision. When you are done, Bango generates a PRISMA flow diagram and a summary of your included literature.

dashboard

Project Dashboard

Your project overview - see how many articles you have and track your progress at a glance.

The dashboard shows you a summary of where things stand: how many articles are waiting to be reviewed, how many have been included or excluded, and any recent activity.
From here you can jump to any part of the workflow using the sidebar or the quick action buttons.
Think of it as your home base - you can always come back here to see the big picture.

rule

Define Your Criteria

Tell Bango exactly what you are looking for by writing your inclusion and exclusion rules.

Before importing any articles, you define the rules for your review. These are the inclusion and exclusion criteria you would normally write in your review protocol.
Each criterion can be given a priority level - from "Critical" (must be met) down to "Optional" (nice to have). This helps the AI make better decisions when rules conflict.
You can also write your research aims here, so the AI understands the broader context of your study.

upload_file

Import Your Search Results

Bring in your search results from databases like PubMed, Scopus, Web of Science, or any source that exports RIS files.

Most academic databases let you export your search results as an RIS file. Bango reads these files and pulls in the title, abstract, authors, journal, year, and keywords for each article.
You can import multiple files - for example, one from PubMed and one from Scopus - and Bango will combine them into a single project.
After import, Bango automatically checks for duplicate records (the same article found in more than one database) and flags them for your review.

science

Review Duplicates

Review and resolve duplicate records that came from searching multiple databases.

When the same article appears in more than one database, Bango detects it and flags it as a potential duplicate.
Exact matches are merged automatically. For close-but-not-identical matches, you can view them side by side and decide whether to keep or remove them.
This step ensures each article only appears once in your review, which is required for accurate reporting.

sell

Review Tags & Labels

Review the categories the AI will use to organize and classify your articles.

Bango suggests content tags (like "clinical-trial" or "qualitative-study") based on the keywords in your articles and the criteria you defined.
It also creates workflow labels (like "priority-read" or "disputed") that help you track the status of individual articles.
You can review, add, edit, or remove any of these before the AI starts screening. This gives you full control over how articles are categorized.

analytics

AI Screening

Let the AI read each article's abstract and decide whether it meets your inclusion criteria.

Once your criteria and tags are set, you start the AI screening. The AI reads the title and abstract of each article and evaluates it against your rules.
For each article, the AI provides a decision (include or exclude), a written explanation, which criteria it matched, suggested tags, and a confidence score.
You can watch the progress in real time. If the AI encounters a problem (like a rate limit from the server), it will retry automatically.
You will need to configure an AI connection in Settings before screening. Bango supports several providers - both cloud-based and locally run models.

description

Review Articles

Browse all articles, read abstracts, and override any AI decisions you disagree with.

After screening, every article has a status: Included, Excluded, or still in the Working list. You can browse, search, and filter to find specific articles.
You have the final say - if you think the AI made a mistake, you can change any decision with one click. All changes are logged in the audit trail.
Use the search bar to find articles by title or keyword. Filter by status, tags, confidence score, or year to focus on what matters most.

account_tree

PRISMA Flow Diagram

Generate a PRISMA 2020 flow diagram for your review report, with all record counts filled in automatically.

The PRISMA flow diagram is a standard requirement for systematic reviews. It shows how many records were identified, screened, included, and excluded at each stage.
Bango generates this diagram for you automatically, with accurate counts drawn from your actual data.
You can choose to show a breakdown of exclusion reasons. The diagram can be exported as an image (SVG or PNG) for inclusion in your manuscript.

summarize

AI Summary

View a synthesis of your included articles - key themes, research trends, methodological patterns, and gaps in the literature.

Once you have a final set of included articles, Bango can produce a structured summary of the body of literature you have identified.
The summary covers recurring themes, common research methods, strengths and weaknesses across studies, and areas where further research is needed.
This can serve as a starting point for writing the narrative synthesis or discussion section of your review.

info

About Bango

Developed by BonCode (Bilal Soylu) with permission from Startup Strategy Advisors LLC. Released as open source under the Apache License 2.0.

bug_report Report Issues & Get Support

Bibliometrics

Bibliometrics is the statistical analysis of books, articles, and other scholarly publications. It helps researchers map scientific fields, trace citation pathways, quantify author impact, and visualize collaborations across institutions and research groups.

Bango includes six interconnected bibliometric modules. They all operate on your project's included articles and the citation/reference data you have imported (see References & Citations). Before any module can display results, you must have citation and references data. Click Refresh from the Bibliometrics dashboard once to build the analytical data layer.

Getting Started

Open Bibliometrics in the sidebar and click the Refresh button. This single transaction parses authors, institutions, terms, and citation links from your active metadata. After it completes, all six analysis modules become available from the dashboard tiles.

What it measures

A co-authorship network represents researchers as nodes and their joint publications as edges. Two authors who have published together get a link, and the more papers they co-author, the stronger that link becomes. The resulting map reveals research teams, institutional clusters, and the central figures who bridge different groups.

How Bango builds it

Absolute edge weights: The link between two authors is the raw count of papers they co-authored. This preserves the real-world intensity of collaboration.
Full & fractional modes: Full counting credits every co-author equally. Fractional counting divides each co-authorship by the number of authors on the paper, reducing the dominance of large consortium papers.
Louvain community detection: Groups authors into color-coded research clusters by optimizing network modularity.
ForceAtlas2 layout: A force-directed simulation pulls tightly-connected authors together and pushes separate groups apart, making collaborative cohorts visually obvious.

Key controls

Modularity resolution: Raise it to split large clusters; lower it to merge them.
Min. articles filter: Hide authors with fewer than N publications to reduce clutter.
Color mode: Color nodes by cluster, or temporally by average publication year.
Layout: Fixed (preserves positions across filters) or Dynamic (re-simulate).
Search & locate: Find a specific author and highlight their neighborhood.

Use cases

Identify the core research group around a prolific principal investigator.
Find isolated sub-fields that rarely collaborate with the mainstream of your corpus.
Spot potential collaborators or reviewers by tracing who bridges two communities.
Support a "social network analysis" section in a bibliometric review paper.

What it measures

Unlike co-authorship, a citation network captures the flow of ideas. Each article is a node, and a directed edge from A to B means "A cites B." This reveals foundational (heavily-cited) works, recent derivatives, and the chronological lineage of a research thread. Main-path analysis then extracts the backbone - the chain of papers that forms the intellectual spine of the field.

How Bango builds it

Directed edges: Drawn from the citing article to each article it references (backward links). Forward citations (cited-by) extend the picture when imported.
Unmatched leaf nodes: Reference papers that are not themselves included articles appear as small dashed grey leaves, so you can see the full intellectual context without polluting the main graph.
Main Path (SPC): Search Path Count highlights the sequence of citations carrying the highest traversal count, dimming everything off the backbone.
Ancestry / Progeny isolation: Select a node to trace its full citation ancestry (what it builds on) or progeny (what builds on it).

Key controls

Min. citations received: Hide weakly-connected nodes.
Show isolated papers: Toggle nodes with zero citation links.
Show unmatched references: Include/exclude the grey dashed leaf nodes.
Main Paths (SPC): Toggle the backbone highlight; off-path nodes dim.
Time-slice: Restrict the graph to a year range via a dual-thumb slider.
Color mode: Cluster or temporal (publication year gradient).

Use cases

Identify the seminal foundational papers that everyone in the field cites.
Trace how a method or concept propagated from its origin to current work.
Find recent breakthrough articles sitting at the tips of citation chains.
Justify the selection of "key papers" in a narrative literature review.

What it measures

A keyword co-occurrence network treats terms as nodes. Two terms are linked when they appear in the same article, and the more articles they share, the stronger the link. This reveals the conceptual structure of a field: which topics are studied together, where the field is fragmenting into sub-specialties, and which terms are central hubs bridging different research themes.

How Bango builds it

Bango is unusual in that it can draw keywords from five different sources, which you can combine or filter independently:

Metadata - author-supplied keywords from the original database records.
AI Noun Phrases - noun phrases extracted from abstracts by an LLM during screening, capturing concepts the authors never tagged explicitly.
Tags - the user-defined content tags assigned during screening.
Labels - screening workflow labels (e.g. "disputed", "priority-read").
User Added - custom keywords you add manually.

Co-occurrence strength equals the number of articles in which two terms appear together. Louvain clustering then groups terms into thematic communities, and ForceAtlas2 lays them out.

Key controls

Keyword sources: Toggle any combination of the five sources (at least one required).
Min. document frequency: Drop terms appearing in fewer than N articles (removes noise).
Min. co-occurrence strength: Hide weak links below a threshold.
Color mode: Cluster or temporal (average year of articles using the term).
Cluster legend: Click a cluster to isolate its thematic neighborhood.

Use cases

Map the thematic landscape of a research area for an introduction or scoping review.
Detect emerging terminology clusters that signal a new sub-field.
Compare author keywords vs. AI-extracted concepts to find blind spots in indexing.
Identify "bridge terms" that connect otherwise separate research communities.

What it measures

The timeline view plots three year-by-year distributions as stacked bar charts: publications (the years your included articles were published), references (the years of the papers they cite), and citations received (the years of papers that cite them, when forward citations are imported). A growth-rate sparkline shows year-over-year acceleration of your own corpus.

How Bango builds it

Publication years come from each included article's publication_year.
Reference years come from the imported backward-reference papers (typically older, showing the intellectual foundation).
Citation years come from forward-citation imports (typically newer, showing current impact).
A secondary Top Journals chart auto-hides below 700px viewport height to keep the primary chart readable.

Key controls & export

Year-range filter: Drill into a specific period.
CSV export: Download the raw year/count data for your own analysis or plotting.
SVG export: Save the chart as a vector image for manuscript figures.

Use cases

Visualize the growth (or decline) of a research field over decades.
Show the gap between the age of your corpus and the age of its references - evidence of a field building on classic vs. contemporary work.
Identify "breakout years" that correlate with landmark publications.
Produce a publication-trend figure for a bibliometric results section.

What it measures

This view moves beyond network structure to quantify individual author impact. Every author in your normalized corpus is ranked on a sortable table. Bango computes three complementary bibliometric indices, all scoped to the included articles you have screened in:

h-index - an author has index h if h of their papers have at least h citations each. Balances productivity and impact.
i10-index - the number of papers with at least 10 citations. A Google Scholar-style productivity measure.
g-index - gives more weight to highly-cited papers than h-index; the top g papers together receive g² citations.
Plus first-author, last-author, and solo publication counts to distinguish leaders from contributors.

How Bango builds it

All metrics are computed locally in Rust from the normalized author tables and are scoped to your included articles (not the global citation count, which may be incomplete). Citation counts draw on the num_cited field parsed from the N1 notes during import.

Key controls

Sortable columns: Click any header (papers, h-index, citations, year) to re-rank.
Detail slide-over: Click an author for a per-article breakdown.
Google Scholar lookup: External-link icons launch a Scholar search for the author's global profile (useful for cross-checking against worldwide metrics).

Use cases

Identify the most influential authors in your corpus for targeted expert outreach.
Compare h-index vs. first-author counts to find emerging vs. established leaders.
Build an "author prominence" table for a bibliometric results section.
Find potential collaborators with complementary productivity profiles.

What it measures

Two papers are co-cited when a third paper cites both of them. The more often they are co-cited, the more the academic community treats them as related. Co-citation analysis therefore reveals intellectual structure from the outside - based on how others group works together - rather than from the authors' own keywords or collaborations. It is especially powerful for mapping the foundational literature of a field.

How Bango builds it

Unlike the other networks which are precomputed during normalization, co-citation is calculated on demand because the result depends on the normalization mode and scope you choose. Bango scans the reference lists of your selected articles, counts shared reference papers, and builds a weighted similarity matrix between citing articles. Four normalization modes transform the raw co-citation counts:

Raw - the absolute number of shared references. Preserves true intensity but favors prolific citing articles.
Cosine - raw count divided by the geometric mean of each pair's total references. The default; balances intensity and size.
Jaccard - shared references divided by the union of all references. Penalizes pairs with many non-overlapping references.
Pearson - correlation coefficient across the reference vectors. Emphasizes structural similarity regardless of raw frequency.

Key controls

Scope: Compute over included articles only, or all articles in the project.
Normalization: Toggle Raw / Cosine / Jaccard / Pearson (re-computes instantly).
Min. citation count: Drop articles cited fewer than N times.
Min. co-citation: Hide edges below a similarity threshold.
Dual visualization: An interactive network graph and a sortable co-citation heatmap of the strongest pairs.
Color mode: Cluster or temporal (publication year).

Use cases

Identify the "invisible colleges" - groups of papers the field consistently reads together even without formal collaboration.
Compare normalization modes: Cosine for a balanced view, Pearson for pure structural similarity, Raw to see dominant citing hubs.
Discover foundational reference pairs that anchor an entire research stream.
Produce a co-citation map figure for a bibliometric review, then cite the top pairs as evidence of thematic clustering.

If you run the same RIS bibliography dataset through both Bango and VOSviewer, you will likely notice differences in the number of clusters and the relative spacing of authors on the screen. This is due to different core mathematical formulations:

Feature / Dimension	Bango Approach	VOSviewer Approach
Link Normalization	Uses Absolute Weights (the exact number of co-authored papers).	Uses Association Strength Normalization (link weight is divided by the product of both authors' total publications).
Modularity & Clustering	Optimizes standard Louvain modularity. Connected groups are kept together into cohesive communities (typically yielding 3-4 clusters on small datasets).	Optimizes normalized modularity, heavily penalizing weakly-associated links. Splits nodes into many fine-grained clusters (typically 14+ clusters on the same small dataset).
Visual Layout	ForceAtlas2 force-directed layout (spring/magnet simulation). Highly connected hubs are drawn strongly to the center.	VOS layout (multidimensional scaling / similarity distance minimization). Relies strictly on normalized association strengths.
Advantages	Reveals real-world collaboration hubs and direct working cohorts. Very easy to read and understand on small-to-medium datasets. Perfect for departmental or local-group collaboration reviews.	Prevents highly prolific authors from dominating the entire map. Allows small, specialized sub-fields to stand out clearly. Very clean cluster boundaries on huge datasets (e.g., 5,000+ papers).
Disadvantages	In large-scale networks, nodes can collapse into a single dense central cluster ("hairball"), requiring manual adjustment of the resolution slider to split them.	Normalization can distort the visual sense of scale. A single paper by two isolated authors can look just as strong as 10 papers by productive authors.

Tip: Neither approach is "wrong." Bango is optimized for interpreting small-to-medium review corpora (the typical 50-2,000 article systematic review), while VOSviewer shines on macro-scale bibliographic datasets. For the same dataset, use Bango's modularity resolution slider to control cluster granularity.

Common Errors & Solutions

If you are seeing an error while using Bango, check the list below. Each entry shows the error message you might encounter, what causes it, and what you can do to fix it.

Cause: Google blocks Gemini API access from the IP address location making the request. Google restricts API availability in certain regions for regulatory reasons. The FAILED_PRECONDITION status confirms your environment fails to meet service requirements. Common causes: the device is in an unsupported country, a VPN routes traffic through a restricted region, or cloud infrastructure (AWS, GCP) operates in a restricted data center zone.

Solution: Ensure the machine executing the API call uses an IP address from a supported region. Verify the current list of allowed countries in the Google Gemini API documentation. Options: upgrade to a paid Gemini API key (pay-as-you-go), connect through a VPN in a supported region, move cloud workloads to a supported zone, or switch to a different provider such as OpenAI or Anthropic.

Cause: You have exceeded the number of requests allowed per minute by your API plan. Free tiers have very low limits.

Solution: Reduce the concurrency setting and increase the request delay in Settings. Bango automatically retries with exponential backoff, but lowering throughput helps avoid hitting limits entirely. Alternatly, you can purchase higher limits with your provider. Please note some provider (Z.AI) also issue this message when you are using the wrong endpoint (Base URL). Please ensure you use the correct Base URL for your subscription. Bango prefills the most common one, yours might be different.

Cause: The API key is invalid, malformed, or has been revoked. Google Gemini returns HTTP 400 instead of the more common 401 for invalid keys.

Solution: Go to Settings and verify your API key is correct. For Google Gemini, ensure the key was generated from Google AI Studio and has not expired. If you recently regenerated the key, paste the new one. This error can also occur if the key belongs to a different Google Cloud project or service.

Cause: The API key is missing, incorrect, revoked, or does not have permission for the requested model.

Solution: Go to Settings and verify your API key is correct. Also check with your provider whether this key is still active. If you recently regenerated the key, paste the new one. Check that your account has access to the model you selected.

Cause: The local AI server is not running or the endpoint URL is wrong.

Solution: Make sure the local server is started (e.g., run ollama serve). Verify the endpoint URL and port in Settings match your server configuration. Check that no firewall is blocking the connection.

Cause: The model name in your configuration does not match any available model on the provider.

Solution: Check the model name for typos. Use the model picker in Settings to see available models for your provider. Model names change over time - make sure you are using the current identifier. For local AI check your tools documentation on how to provide the name and where you can look it up.

Cause: The combined input (criteria + article text) and expected output exceed the model's maximum context window size.

Solution: Reduce the batch size in Settings, shorten your criteria text, or switch to a model with a larger context window (e.g., 128K or 200K token models).

Cause: The AI returned an invalid or unexpected response format. This is more common with smaller or locally-run models.

Solution: Retry the individual article. If the problem persists, try a different model or a larger quantization. Cloud providers (GPT-4, Claude) produce more reliable structured output.

Cause: No API key has been entered in the Settings screen.

Solution: Go to Settings, select your provider, and enter a valid API key. Click Save to persist the configuration.

Cause: The endpoint is using a self-signed or expired SSL certificate, which the app cannot verify.

Solution: For local providers, use http://localhost instead of https://localhost. For remote self-hosted servers, ensure a valid certificate is installed.

Cause: Local models run on your hardware and may be slow, especially with large context windows or limited RAM/VRAM.

Solution: Reduce the context window setting, use a smaller quantized model (Q4_K_M), lower concurrency to 1, and close other applications to free memory.

Cause: The model requires more RAM or VRAM than is available on your system.

Solution: Use a smaller model or a more aggressive quantization (Q4_K_M instead of Q8). Close other applications. For GPU-based inference, ensure your GPU has enough dedicated VRAM.

info

Still stuck?

If your issue is not listed above, please open an issue on GitHub and the community will help you out. Include the error message, your provider, and what you were doing when it happened.

bug_report Report Issues & Get Support

Local & Free AI

Bango can use either a cloud LLM provider or a model running on your own hardware. This page covers both options: free cloud API keys you can get in minutes, and a complete guide to running models locally for full privacy and zero per-request cost.

Rate-limited access to the Gemini family (including Gemini 1.5 / 2.0 Flash and Pro) through Google AI Studio. The most generous ongoing free tier available, with very large context windows well-suited to Bango.

Go to Google AI Studio and sign in with a Google account.
In the top bar (or left sidebar) click “Get API key” → “Create API key”.
Copy the generated key (it starts with AIza...).
Free-tier requests may be used by Google to improve products; switch to a paid billing plan if you need a no-training data guarantee.

model_training Recommended for Bango: gemini-1.5-flash or gemini-2.0-flash (fast, cheap, large context). Use gemini-1.5-pro for harder reasoning at lower throughput.

open_in_new aistudio.google.com

The Experiment tier gives rate-limited access to all Mistral models, including the open-weight and “open-mistral-” series. A solid always-on option with no expiry.

Create an account at console.mistral.ai (La Plateforme).
Navigate to API Keys and generate a new key.
Copy the key.
The Experiment tier is intended for prototyping; throughput (requests per minute) is capped and may change over time.

model_training Recommended for Bango: open-mistral-nemo or mistral-small-latest. Both have a context window large enough for abstract screening.

open_in_new console.mistral.ai

Z.AI offers free access to specific lightweight GLM models (such as GLM-4-Flash / GLM-4.5-Flash) via an OpenAI-compatible API. A good choice for cost-free screening runs.

Sign up at z.ai and verify your account.
Open the API / API Keys section of the console and create a key.
Copy the key.
In Bango, choose the built-in “z.ai” provider; only the whitelisted free models are complimentary.

model_training Recommended for Bango: glm-4-flash or glm-4.5-flash (free). Confirm the current free-model list in the Z.AI console before large runs.

open_in_new z.ai

New OpenAI accounts receive a small one-time credit (about $5) that is valid for roughly three months. Useful for trying Bango end-to-end, but not a long-term free option.

Create an account at platform.openai.com and add a phone number for verification.
The sign-up credit appears on the Billing → Usage page.
Generate an API key under API Keys → Create new secret key.
Once the credit is spent or expires, you are billed per token. Set a monthly spending limit if you keep the key active.

model_training Recommended for Bango: gpt-4o-mini (cheapest). Use gpt-4o for higher-quality reasoning when budget allows.

open_in_new platform.openai.com

New Anthropic console accounts occasionally receive a small promotional credit (about $5) valid for a limited time. Handy for an initial screening run, after which billing applies.

Create an account at console.anthropic.com and verify your phone number.
Check the Billing / Credits section for any promotional balance.
Generate an API key under API Keys → Create Key.
Promotional credit availability and expiry vary by region and over time; confirm on the Billing page.

model_training Recommended for Bango: claude-3-5-haiku-latest (cheapest, fast). Use a claude-sonnet variant for stronger reasoning.

open_in_new console.anthropic.com

Provider	Ongoing Free Tier	Included Access	Expiration
Google	Yes	Rate-limited access to Gemini models via AI Studio	Never
Mistral AI	Yes	Rate-limited access to all models via Experiment tier	Never
Z.AI	Yes	Access to specific lightweight models (e.g., GLM-4.5-Flash)	Never
OpenAI	No	One-time $5 sign-up credit	~3 months
Anthropic	No	One-time $5 sign-up credit	Limited time

settings

Use a free key in Bango

Open Settings from the sidebar.
Under AI Provider, select the matching provider (e.g. “Google Gemini” or “Mistral AI”).
Paste your API key and (optionally) pick a model from the list.
Click Test Connection. If it succeeds, you are ready to run AI screening, summaries, and chat.

Running models on your own hardware means your data never leaves your machine - full privacy, no API fees, offline capability, and predictable performance. Tools like Ollama and llama.cpp make it straightforward to deploy models on laptops, workstations, or servers.

Bango sends article abstracts and criteria together, so models supporting 50K+ token context windows are strongly recommended.

Minimum: 16 GB RAM for small models (e.g., Phi-3 Mini 3.8B at 5-10 tokens/sec).
Recommended: Apple Silicon Mac (M1/M2+) or a PC with a dedicated GPU (8-16 GB VRAM) for 7B+ parameter models.
For 50K+ token context, you typically need 16 GB+ VRAM or unified memory.

Ollama is the easiest way to get started. It handles model management automatically and provides a simple interface.

Install Ollama.
curl -fsSL https://ollama.com/install.sh | sh
Pull a model.
ollama pull llama3
Start the server.
ollama serve
In Bango Settings, set the provider to "Ollama" and the endpoint URL to:
http://localhost:11434/v1

lightbulbOn Windows, download the installer from ollama.com instead of using curl.

LM Studio provides a graphical interface for downloading and running models.

Download LM Studio from lmstudio.ai and install it.
Browse and download a model (e.g., Llama 3 8B Instruct).
Go to the Local Server tab and click Start Server.
In Bango Settings, set the provider to "LM Studio" and the endpoint URL to:
http://localhost:1234/v1

llama.cpp offers maximum flexibility and performance with advanced quantization options. It runs on a wide variety of hardware and gives you fine-grained control over inference parameters.

Build from source or download a pre-built binary from github.com/ggerganov/llama.cpp.
Run the server with your model.
./llama-server -m model.gguf --port 8080
In Bango Settings, set the provider to "llama.cpp" and the endpoint URL to:
http://localhost:8080/v1

Choose models that support 50K+ token context for Bango screening. Good options: Llama 3 (8B or 70B), Mistral (7B), Qwen 2.5 (7B or 14B), and Phi-3 Medium.

lightbulbUse quantized versions (Q4_K_M or Q5_K_M) for lower memory usage with minimal quality loss. Quantization reduces model precision (16-bit to 4-bit) to shrink memory requirements - Q4_K_M is roughly 40% of the original size and Q5_K_M about 50%.

Temperature: 0.2 (default, good for analytical tasks).
Concurrency: 1 (local models handle one request at a time).
Request delay: 1000ms+ to give the model time between requests.
Context window: set this to match your model's actual capability - do not set it higher than what the model supports.
If you get truncated responses, lower the context window or switch to a model with a larger window.

info

Feature Reference

A detailed reference for every screen and feature in Bango. Jump to any section below.

Bango is designed with a specific philosophy: AI should serve as a high-accuracy, transparent research assistant that helps you organize and analyze literature, not a black-box decision-maker that replaces human oversight.

How Bango Works with AI

Traditional AI chat interfaces require you to copy-paste or upload entire PDFs into a single context window. In contrast, Bango treats papers the way a human researcher does: section by section, with the most relevant passages surfaced on demand.

Section-Aware PDF Extraction: When you attach a PDF, Bango parses the text and splits it into structured Method, Results, and Discussion passages. This ensures important sections are not lost to context window limits.
Criteria-Targeted Chunking: Rather than feeding an entire paper to the AI, Bango's retrieval engine ranks and pulls only the top-K passages matching your criteria.
Grounded Citation Mapping: Every AI summary, screening advice, and chat answer is directly mapped to the source passage (e.g., Smith 2023 (§Methods)). These labels are interactive links that open the source text, so you can verify the AI's work.
Deterministic Logic Layer: Bango evaluates conflict resolution rules locally in SQLite based on your criteria priorities, using the LLM solely to advise on matches.

Bango RAG vs. Dumping Documents into Chat Windows

Dimension	Dumping in Chat Windows (ChatGPT/Claude UI)	Bango Structured Retrieval (RAG)
Context Truncation	High. Long papers exceed limits and get silently cut off.	None. Section-by-section parsing and word budgets prevent context overflow.
Hallucination Risk	High. The model generates summaries based on incomplete or recalled memory.	Low. The model is constrained to analyze only target, extracted passages.
Verifiability	Low. Answers are plain text; you must search the PDF to locate the source.	High. Interactive passage links (e.g., `§Methods`) open the exact source text.
Reproducibility	None. Repeated questions yield different outputs; there is no saved history.	Full. Every decision is logged in an immutable database audit trail.
Structured Comparison	Manual. You must copy/paste data to spreadsheets to compare papers side-by-side.	Automatic. Standard variables (sample size, design, effect sizes) are parsed into cards.
Setup Complexity	Instant. Just upload or paste text and type.	Structured. Requires importing bibliography files and attaching PDFs.

Benefits & Drawbacks of Bango's Approach

While Bango's structured approach provides the rigor needed for systematic reviews, it represents a different set of trade-offs:

Benefits: Perfect audit trails for PRISMA reporting, automatic re-evaluation when criteria change, and lower API token costs by avoiding sending unnecessary parts of the paper.
Drawbacks: Processing requires a multi-phase pipeline (import, text extraction, chunking, and database mapping) that takes longer to run than a simple chat prompt, and requires attaching original PDF files.

The Project Dashboard is the home base of your systematic literature review. It provides a real-time status summary of your article library and screening progress.

Key Features

Progress KPIs: Cards displaying the counts of articles in different states: Working (unscreened), Included, Rejected, and unresolved Duplicates.
Recent Activity Trail: An active chronological log showing the latest system and user actions, such as imports, status overrides, and AI screening updates.
Quick Actions: Navigation buttons to jump directly to key screening tasks.

Before starting your AI screening, you must define the boundaries of your review. The Criteria Editor allows you to formulate research aims and explicit screening rules.

Key Features

Research Aims: Free-text statements defining the broader research context. These are sent as prompts to the LLM to guide screening decisions.
Inclusion / Exclusion Criteria: Explicit conditions articles must meet (Inclusion) or must not meet (Exclusion).
Priority Conflict Resolution: Each criterion must be assigned a priority (Critical, High, Standard, Low, or Optional). If both inclusion and exclusion criteria match:
1. The highest-priority matching criterion wins the decision.
2. If priorities are tied, inclusion wins.
3. If no criteria match at all, the article is excluded.

Screen Controls & Buttons

Add Aim / Add Criterion: Buttons that open input forms to add items to the active list.
Priority Selector: Dropdown menus beside criteria to update priority values.
Edit / Delete: Actions on each row to modify text or permanently delete criteria.

Bango supports importing bibliography results exported from academic search engines.

Key Features

Multi-Format Support: Reads both standard RIS (.ris) and BibTeX (.bib) files.
Capacity Guard: The system enforces a project capacity limit of 10,000 articles. Imports that would exceed this threshold are blocked.
Exclusion Filters: A preview list of imported papers allows you to review metadata and manually deselect individual rows before writing them to the database.
Metadata Validation: Articles missing essential fields (title, abstract text, or authors) will raise warnings and be skipped during import.

Format Examples

RIS Format:

TY  - JOUR
TI  - AI-Assisted Systematic Review Abstraction
AU  - Soylu, Bilal
JO  - Journal of Advanced Agentic Coding
PY  - 2026
AB  - This paper describes a novel system for abstract screening...
KW  - systematic review
KW  - LLM
ER  -

BibTeX Format:

@article{soylu2026ai,
  title = {AI-Assisted Systematic Review Abstraction},
  author = {Soylu, Bilal},
  journal = {Journal of Advanced Agentic Coding},
  year = {2026},
  abstract = {This paper describes a novel system...},
  keywords = {systematic review, LLM}
}

When combining bibliography files from multiple databases, duplicate entries are common. Bango runs a multi-strategy deduplication pipeline.

Deduplication Strategies

DOI Exact: Matches DOI strings exactly. Auto-merges duplicates.
Title + Year Exact: Matches normalized title (similarity >= 95% via Levenshtein distance) and exact year. Auto-merges.
Fuzzy Title + Year: Matches title with 70-94% similarity and exact year. Flags for manual resolution.
Author + Title Partial: Matches first author's last name exactly and normalized title similarity >= 80%. Flags for manual resolution.

Screen Controls & Buttons

Merge: Resolves the conflict by linking the duplicate record to the parent article and marking it as read-only.
Keep Both: Dismisses the duplicate flag, keeping both records in the active working set.
Skip: Defers decision to review later.

Categorize and track your literature using tags and labels.

Tags vs. Labels

Tags (Content categories): Used to classify the subject matter of an article (e.g., "randomized-control", "neural-network"). Suggested by AI during screening or created manually by users.
Labels (Workflow markers): Used to track organization or audit status (e.g., "disputed", "priority-read"). Generated by AI or managed manually.

Screen Controls & Buttons

Color Swatch: Click color buttons to assign specific colors to your tags and labels.
Create Tag/Label: Input field at the top of the panels to add a new category directly to the project vocabulary.

The AI Screening view manages running your screening queue using remote or local LLM models.

Required Checklist

Before the screening worker can start, the system checks for:

At least one Research Aim.
At least one Inclusion Criterion and one Exclusion Criterion.
A valid LLM Provider configuration and API key in Settings.

Behaviors & Exception Handling

Token Limit Safeguard: The system estimates input sizes for each abstract batch and prompts warnings if the estimated payload exceeds 80% of the configured context window (requires at least 50,000 tokens context).
Rate Limiting (HTTP 429): If the remote provider throttles the connection, Bango pauses and executes automatic retries using exponential backoff. You can mitigate this by lowering concurrency or increasing request delays in Settings.
Authentication Failure (HTTP 400/401): Occurs if API keys are invalid or revoked. Check the Diagnostics log and verify configurations in Settings.
Screening Errors: API timeouts or malformed JSON responses leave the article in Working status with the "Screening Error" flag enabled. You can retry these individual records from the detail panel.

Screen Controls & Buttons

Start Screening: Launches the background async queue worker.
Pause / Resume: Safely halts the queue execution or resumes it from where it stopped.

Bango can translate non-English articles to English before AI workflows process them. This is a permanent rewrite: after translation, the working article fields hold English text, and the originals are preserved in archive tables but currently not visible. Translation is experimental and opt-in. To enable it use the Auto Translate setting in the Settings page.

How It Works

When Auto Translate is enabled in Settings, Bango detects the article's language from the RIS/BibTeX language field. English articles and articles with blank language are skipped. Non-English articles are translated using your configured LLM provider.

The scope of translation is chosen automatically from whether the article has full text attached:

MetadataOnly: Translates the title and abstract. Used during import when no full text is attached, and during the screening pre-translation step.
FullText: Translates the title, abstract, all article chunks, then re-chunks the stitched English result for screening. Used when full text is attached and during batch import.

Implicit Translations

Please note: Even if you do not have Auto Translate enabled you may be causing implicit translations via the LLM that you use. When you submit non-english articles for AI screening or AI Summary the LLM you choose decides whether to translate these and or respond in english or the original language. Thus explicit translations via enabling the Auto Translate setting may be preferable as they will be more reproducable.

Translation State Machine

Each article tracks its translation lifecycle through five states which you can also review in the audit logs for articles:

none : Never translated. Initial state.
queued : Job sent to the worker. Waiting to be processed.
running : Bango is actively translating (LLM call in flight).
succeeded : Translation complete - with a time stamp.
failed : Translation errored. This holds the error message. You can retry via the translate button.

Screening-Time Translation

When Auto Translate is on, the screening engine explicitly runs a pre-translation step before the AI reads any abstracts. It queries unscreened working articles with non-English language, enqueues Metadata (abstract and keywords) translation jobs, and waits for all to complete. The screening progress bar shows "Translating 3/12 articles..." during this stage. The readiness check also reports pending so you know how many articles need translation before screening can begin.

Manual Translation

A translate button appears on the article detail header for any non-English article. Clicking it opens a confirmation dialog (warning about the permanent rewrite and token cost), then enqueues the job. Popup messages provide feedback at each stage. A red TRANSLATED badge appears on the header once translation succeeds.

Original Content Archive

Before the working articles row is rewritten, the original-language text is saved to two dedicated tables. These are currently not visible in UI but can be queried from database with other tools:

article_original_content : Stores original title, abstract, full text, and the source language at translation time.
article_original_chunks : Stores the pre-translation chunk coordinate space (section, content, word count). After translation, English chunks live in article_chunks with their own independent indices.

Batch Import Integration

Translations are also integrated into batch imports and imported full text pdf will also automatically translated if the setting is enabled.

Multilingual Section Classification

To correctly chunk non-English full text, Bango's section classifier supports 10 languages beyond English: French, Spanish, Japanese, Chinese, German, Russian, Portuguese, Italian, Arabic, and Turkish. Each language maps academic section keywords (Abstract, Introduction, Methods, Results, Discussion, Conclusion, References) to its native terms. A Unicode-aware numbered-heading regex detects headings in non-Latin scripts (Cyrillic, CJK, Arabic).

The article viewer uses a dense 3-pane layout allowing you to read abstracts, inspect metadata, examine AI reasoning, and download attachments.

Key Features

Left Pane (Filters & Search): Query by title, year, tags, AI confidence, or status.
Center Pane (Article Table): Scroll through matched articles and status indicators.
Right Pane (Sliding Details Panel):
- AI Decision Card: Shows the suggested action, matching criteria list, and confidence.
- Full-Text Attachments: Attach PDFs or TXT files. Raw texts are extracted and cached locally in the full-text storage directory.
- Inline PDF Reader: Render and read attached documents side-by-side with the metadata.
- Audit Trail Timeline: View immutable histories of actions taken on the article (imports, status changes, manual overrides, error codes).

Screen Controls & Buttons

Include / Reject: Action buttons at the top of the detail panel that override AI-assigned decisions. These are logged in the audit trail.
Attach File: File selection trigger that imports a PDF or text file for full-text caching.

Bango tracks both backward references (articles cited by the paper) and forward citations (articles that cite the paper) for included records.

Import References Pop Up

To load specific reference datasets for a given article, click the Import button on the References tab in the article details panel. This opens the Import References Dialog where you can:

Select the citation direction: Backward (cited refs) or Forward (cited by).
Click Choose File to upload an RIS or BibTeX file matching that citation dataset.
Review the parsed preview list, and click Add to import them.
Click the Help link in the popup header to close the dialog and jump back to this reference chapter.

Data Sources

Citation count fields (num_cited and num_references) are automatically parsed from the N1 field of bibliography files during initial imports. Detailed citation lists can be obtained from:

Web of Science: Export records selecting "Full Record and Cited References" in RIS format.
Lens.org: Create a free account on lens.org, compile a collection, and export the citation/reference lists in RIS or BibTeX format.

Match Status States

Once references are loaded, each reference paper has one of the following states:

unmatched: The paper is listed in the reference dataset but has not been found or promoted in the library.
matched: The paper matches an existing article in the main library (matched by DOI or title/author/year). Clicking the link icon will take you to that library record.
imported: The paper has been promoted to a full article and is now available in the working list.
not_in_library: No match exists in the current library.

The Bibliometrics tab analyzes structural collaborations, citation density, and keywords across six modules: Co-Authorship, Citation Network, Keyword Co-Occurrence, Publication Timeline, Author Productivity, and Co-Citation Analysis. See the Bibliometrics help tab for detailed explanations and use cases for each module.

Modularity & Layout Algorithms

Bango uses a local modularity engine designed to map direct working relationships:

Feature	Bango Engine
Normalization	Uses Absolute Link Weights (actual co-authored papers).
Clustering	Standard Louvain modularity optimization. Groups nodes into highly cohesive communities.
Visual Layout	ForceAtlas2 force-directed model. Highlights central hubs and direct departmental working groups.

Analytical Features

Normalize: Populates the bibliometric databases by parsing active metadata within a single transaction to ensure maximum speed.
Louvain Clustering: Groups co-authors into color-coded collaborative teams. Adjust the modularity slider to group or split cohorts.

The Chat Assistant provides a conversational interface to query your systematic review database.

Key Features

RAG (Retrieval-Augmented Generation): The assistant uses your project's inclusion/exclusion criteria, research aims, and article abstracts as context to answer questions.
Source Citations: Answers include direct citation badges linking to referenced papers. Clicking a badge opens the corresponding article details panel.
Two retrieval modes: the default Article Chat (dumps selected article summaries into the prompt) and the token-optimized Wiki Chat (BM25 retrieval over the wiki index - see the Wiki section for why this scales to hundreds of pages).

Bango Chat vs. Copy-Pasting into a Plain LLM / Obsidian

The defining cost of an LLM query is the input token count - the context you send with every question. Bango's two chat modes are engineered to minimize that cost, and the difference compounds with every follow-up question:

Approach	Tokens per question	Scales to
Bango Article Chat (`send_chat_message`)	~7,500 tokens for 10 articles. Each article's `summary_text` (~3,000 chars / 4 chars/token) is re-sent on every question. Fine for small, targeted selections.	~20 articles before hitting typical 16K-token context limits.
Bango Wiki Chat (`wiki_chat`)	~3,000 tokens, flat. The question is BM25-matched against the SQLite FTS5 index; only the top 3-5 matching pages are retrieved, capped at a `12,000-char` (~3,000-token) budget. Overflow pages are listed as titles only ("see also").	Hundreds of pages. Retrieval picks the relevant subset regardless of corpus size - a 5-page wiki and a 500-page wiki cost roughly the same.
Manual Obsidian + LLM	You copy-paste notes by hand. No retrieval, no budgeting - easy to overflow the context window or pay for tokens you never use.	Unmanageable past ~10 notes per question.

Why the token budget matters

At typical LLM pricing (~$3 / 1M input tokens for mid-tier models), 100 follow-up questions on a 10-article selection costs about $2.25 in Article Chat vs. $0.90 in Wiki Chat - and only the latter keeps working as your corpus grows. The FTS5 index is built once (during wiki ingest) and queried offline in milliseconds; there is no per-question embedding API call.

Screen Controls & Buttons

Article selection (+): choose which articles to include as context for Article Chat.
Wiki toggle: the local_library icon (right of the (+) button) switches to Wiki Chat mode. Visible only when the wiki is initialized and has pages.

The Research Gap Report is an AI-generated analysis that identifies unexplored or under-explored research areas in your included article corpus. It is available from the Summary screen.

What It Does

Reads all included article summaries and key insights to build a comprehensive picture of the current research landscape.
Identifies topics, methods, populations, or angles that are missing or underrepresented across your corpus.
Produces a structured report with gap themes, supporting evidence from existing studies, and suggested research directions.

Use Cases

Framing “future work” sections in your own publications.
Writing grant proposals that target genuine knowledge gaps.
Validating that your systematic review scope is comprehensive.
Identifying niche areas where your research team has a competitive advantage.

How to Use

Navigate to the Summary screen in the sidebar.
Click the Research Gap Report button in the toolbar.
The report is saved in the database and remains available until you regenerate it.

Note: The gap report uses token-optimized retrieval, so even large corpora stay within your LLM's context window budget.

The Search Strategy Builder generates database-ready Boolean search strings from your research aims. It is available on the Criteria screen and uses AI to translate plain-language aims into structured query syntax.

How It Works

The AI reads all research aims from the Criteria screen to understand your review scope.
It produces structured queries with MeSH terms, free-text keywords, Boolean operators (AND, OR, NOT), and field tags (e.g. [Title/Abstract] for PubMed).
The result is displayed as a formatted card you can copy directly into PubMed, Scopus, Web of Science, or any other academic database.

How to Use

Navigate to the Criteria screen in the sidebar.
Ensure you have at least one research aim and a configured LLM provider.
Click the Generate Search Strategy button (identified by the sparkle icon).
The result appears as a card below the button. Copy any query string with one click.

Tip: Regenerate the strategy whenever you update your aims to keep your database queries aligned with your review scope.

The Custom Screening Instructions panel (Section 4 of the Criteria screen) lets you define advanced logic beyond the standard priority-based matching: AND/OR gates, hard exclusions, and conditional inclusion rules that give your AI fine-grained control over screening decisions.

How to Write Rules

Reference criteria by their numbered position shown on the Criteria screen. Inclusion criteria are numbered 1..N and exclusion criteria continue N+1..N+M, so every criterion has a unique global number.
AND gates: “Inclusion criteria 2, 3, and 4 are mandatory AND gates - all three must match for inclusion.”
Conditional inclusion: “Only if criteria 2–4 are all satisfied, consider inclusion criterion 5 OR 6 as the final signal.”
Hard exclusions: “Exclusion criterion 9 is a hard gate; if it matches, ignore inclusion criteria 11–14.”
Combined guarding: “If inclusion 3 and 7 both match, exclusion 5 OR 6 must NOT match for inclusion.”

Click the help icon (?) next to the heading for a live syntax guide with more examples.

Auto-Save & Check Rules

Auto-save: Your instructions save automatically when you leave the text area (click outside) or navigate to another screen. No manual Save button needed.
Check Rules: The sparkle button in the section header runs an AI consistency review over your entire ruleset (aims, criteria, and custom instructions), flagging contradictions, ambiguity, and missing edge cases.
Saved instructions are injected into every screening prompt as a ## Custom Screening Instructions section.

Tip: Leave the field blank for the default priority-only behavior. Rules are only applied when you start a new screening run.

The Wiki is a local-first, Obsidian-style knowledge base that Bango's LLM builds from your included articles. Instead of a flat list of papers, you get a linked, navigable synthesis of concepts, authors, methods, and themes, with every claim traced back to its source article.

Why a Wiki?

A systematic review often ends as a static table. The Wiki turns that table into a living knowledge graph: the LLM extracts entities (sugar tax, SDIL, key authors, methods), cross-links them with [[wikilinks]], and cites each fact with a source reference like [^art-123] that jumps back to the original article. It is the fastest way to understand the landscape of your corpus.

Why Bango Instead of Pure Obsidian?

You can open the generated wiki-root/ folder in Obsidian as a read-only companion view (the Markdown is fully portable). But Bango and a plain Obsidian vault are not interchangeable - Bango's wiki is the output of a review pipeline that Obsidian has no equivalent of:

1. The corpus is already reviewed and normalized

Every page in the Bango wiki is built from articles with status = 'included' - meaning each source has already passed:

Deduplication (4 strategies: DOI exact, title+year ≥ 95%, fuzzy 70-94%, author+title ≥ 80%).
AI screening against your aims/criteria with priority-based conflict resolution.
Your manual review (Include/Reject overrides, full-text attachment, notes).
Bibliometric normalization - the 8-step biblio_normalize pipeline populates canonical author, keyword, and institution tables that the wiki pre-seeds from (see next point).

In Obsidian, you would manually curate which sources are worth synthesizing - there is no deduplicated, screened, normalized corpus to build on.

2. Deterministic 4-layer pre-seed matrix (before the LLM runs)

This is the core advantage. Before the LLM generates a single page, Bango writes a connected graph backbone from the normalized metadata - so the wiki is never missing key structural pages, regardless of which LLM model you use:

Layer	What Bango writes deterministically (no LLM)	Obsidian equivalent
Author pages	One per corpus author, with h-index, total citations, first-author count, papers/year, frequent collaborators as `[[links]]` - all derived from the `biblio_authors` table, not LLM-hallucinated.	You write each by hand, or nothing.
Synthesis pages	One per included article, built from the structured `full_text_ai_summary` JSON (summary + key insights). Slug = article UUID so `[^art-uuid]` citations resolve.	One note per paper, written by hand.
Concept hubs	Top-25 keyword pages, statistically derived from `biblio_terms` co-occurrence - not guessed by the LLM.	You decide what concepts matter, manually.
Source pages	One per uploaded external document (PDF / web / TXT), so external sources get a first-class wiki node and their citations resolve.	Each external doc is an orphan note unless you link it.

The result: a connected graph (author ↔ synthesis ↔ concept ↔ source) exists before the LLM runs. The LLM then enriches it with concept relationships, method groupings, and cross-cutting synthesis - but the backbone is always there.

3. Ingest is batched, parallel, and self-consolidating

Bango's ingest engine is designed to handle large corpora efficiently:

Token-budgeted batching: raw sources are split into batches sized to 40% of your configured context window (the remainder is reserved for the model's output). Example: a 50K-token context → ~80K chars per batch; 20 articles x 2,000 chars = 40,000 chars fits in one batch.
Parallel dispatch: all batches run concurrently via a Bango Orchestrator, bounded by your available LLM Context Size. Each batch carries a compact full-source index so the model can link across batches.
Deterministic consolidation: when multiple batches run, a no-LLM merge pass deduplicates near-identical pages (slug Jaccard similarity ≥ 0.5, or ≥ 2 shared source articles), rewrites inbound links to canonical slugs, and unions the frontmatter. This prevents the childhood-obesity vs obesity-childhood fragmentation that independent LLM calls would otherwise produce.
Cost example: a 50-article corpus on a 128K-token model typically ingests in 2-3 parallel batches (~$0.05-0.10 total at mid-tier pricing). The same corpus processed one-note-at-a-time through an Obsidian plugin would cost 10-50x more in API calls and produce inconsistent cross-linking.

Obsidian's role

Obsidian shines as a reading and annotation surface for the wiki Bango generates - its graph view, backlinks panel, and mobile sync are excellent for exploration. But it cannot produce the wiki: it has no dedup pipeline, no criteria-based screening, no bibliometric normalization, and no batched cross-consolidating LLM ingest. Use Bango to generate and maintain the wiki; use Obsidian (optionally) to read it.

Where Your Documents Live

The Wiki is stored as plain Markdown on your disk, as a sibling of the full-text directory:

~/Documents/Bango/
  fulltext/          # article PDFs and text extracts
  wiki-root/         # the LLM Wiki (plain Markdown)
    AGENTS.md        # the LLM's workflow contract (read on every ingest)
    raw/             # sources: article exports and your dropped files
    wiki/            # generated pages (concepts/ authors/ methods/ synthesis/)
      log.md         # append-only audit trail of ingest and lint runs
    templates/       # page skeletons the LLM follows

Supported Document File Types (Add Documents)

Bango's wiki ingest accepts a wide range of file formats for the Add Documents feature. The raw extraction layer handles each type appropriately so the LLM always receives clean text:

Extension(s)	Source kind	Extraction method
`.txt`, `.md`	Plain text	Direct read (UTF-8).
`.pdf`	PDF	Text-layer extraction (pdf-extract). Scanned-only PDFs produce empty text and are skipped.
`.docx`	Word	DOCX XML body text extraction (docx-rs).
`.html`, `.htm`	Web / HTML	Tag stripping with element-based heuristics.
`.epub`	eBook	EPUB spine document text extraction (epub crate).
`.ris`, `.bib`	Bibliography	Parsed via the same importer used by article import; abstract/title fields become the source text.
`.csv`	Spreadsheet	Row-joined text (headers + cell values).
`.xlsx`	Excel	Sheet-by-sheet text dump (calamine).
`.json`	JSON	Formatted pretty-print of the parsed tree.
`.xml`	XML	Tag-stripped text content.
`.rtf`	Rich Text	RTF control-word stripping.

Files without a recognized extension or unsupported types (e.g. .png, .mp4) are listed in the export report under unsupported and excluded from the LLM prompt.

Export Wiki as Website

The Export Wiki Website option (Wiki toolbar or the global Export menu in the article list) packages your entire wiki as a self-contained static .zip website:

Every wiki page becomes a standalone .html file with full navigation.
Article references resolve to metadata-only stub pages (no full text - copyright safe).
Includes a built-in search engine, styling, and a clickable graph view with ForceAtlas2 layout.
Extract the zip and open index.html in any browser - no server required.

Use this to share your research synthesis with colleagues, submit as supplemental material, or archive your review for posterity.

Hosting Your Wiki Online

The exported wiki is a folder of static HTML files. You can host it for free on any static site hosting service -- no backend, no database, no server configuration needed:

GitHub Pages

Create a GitHub repository, upload the extracted wiki folder, and enable GitHub Pages in the repository settings (Settings → Pages → deploy from main branch).
Your wiki is live at https://your-username.github.io/repo-name/.
Free, fast, and updates automatically when you push new exports.

Netlify

Drag and drop the extracted wiki folder onto Netlify Drop.
Your site is live instantly with a shareable URL. Connect a custom domain for free.

Vercel

Upload the folder to a Git repository or use the Vercel CLI (vercel command in the extracted folder).
Deploys instantly with automatic HTTPS and a shareable URL.

Tip: If you prefer keeping things offline, you can also share the extracted folder via cloud storage (OneDrive, Google Drive). Recipients download and open index.html in their browser - no hosting required.

Configure AI connections, custom directories, reprocessing tasks, and project backups. Bango's settings are arranged into modular cards to help you manage your workspace.

API Key Encryption

To protect your credentials, LLM API keys are encrypted locally using **AES-256-GCM**. The decryption key is derived cryptographically from your local machine's hostname, username, and a secure app salt. API keys are never included in project backups.

Configurable Options & Cards

LLM Provider Settings: Select your AI provider (Google Gemini, Anthropic Claude, OpenAI, Ollama, LM Studio, or custom endpoints) and enter your credentials. Use the model picker to select active models.
AI Summary Settings: Toggle the "Include Section Summaries" option. When enabled, the AI reads your PDFs section-by-section (Methods, Results, Discussion) to build a structured breakdown of study design, sample size, population, effect sizes, and limitations.
AI Screening Preferences: Configure the active screening mode:
- Abstract Mode: Evaluates articles using title and abstract text alone (default).
- Enhanced Mode: Evaluates abstract plus the top criteria-matched passages from full text.
- Two-Stage Mode: Screens abstracts first, then runs a full-text pass only for borderline papers (confidence in configurable range, e.g., [0.4, 0.7)).
Allows setting the chunk budget per article (default 2400 words) and active sections (default Methods, Results).
File Storage: Defines the Bango documents root folder. All cached files, PDF attachments, scrapers, and the local Wiki files reside in subdirectories here (fulltext/, ris/, wiki-root/).
Maintenance & Imports (Reprocessing):
- Rebuild Text Chunks: Forces Bango to re-parse and split attached full-text files into vector chunks.
- Batch Import: A three-phase automated pipeline that scans your Storage directory. It links PDFs to articles via DOI (Phase 1), imports Citation Chaser/RIS metadata (Phase 2), and pre-generates AI summaries (Phase 3).
Project Management: Contains core project options. Export backup as a .bango.json file, import a backup to restore data, reset the current project database, or delete all data.
Diagnostics & Notification History: View previous toast messages, system logs, and error trails.

Bango projects can be exported to a single JSON file and restored on any machine.

Export

Project Backup: Exports all source tables (aims, criteria, articles, tags, labels, audit, references, config) to a .bango.json file.
The 9 biblio_* tables are not exported - they are dynamically generated by biblio_normalize and would bloat backups and trigger constraint violations on import.
Note: The Bango Documents directory (full-text PDFs + Wiki) is not included in the backup. Back it up separately.

Import / Restore

Import a .bango.json file to replace the current project data.
Uses INSERT OR IGNORE + ID-remap maps for tables with UNIQUE constraints to handle older backups.
After import, mark_biblio_needs_refresh ensures the frontend auto-regenerates bibliometric tables.

Delete All Data

Removes all project data, drops all database tables, and deletes the on-disk wiki-root/ directory.
The journal index and app settings are preserved.

Help & Guides

The Big Picture

Project Dashboard

Define Your Criteria

Import Your Search Results

Review Duplicates

Review Tags & Labels

AI Screening

Review Articles

PRISMA Flow Diagram

AI Summary

About Bango

Bibliometrics

Getting Started

What it measures

How Bango builds it

Key controls

Use cases

What it measures

How Bango builds it

Key controls

Use cases

What it measures

How Bango builds it

Key controls

Use cases

What it measures

How Bango builds it

Key controls & export

Use cases

What it measures

How Bango builds it

Key controls

Use cases

What it measures

How Bango builds it

Key controls

Use cases

Common Errors & Solutions

User location is not supported for the API use

Rate limited (HTTP 429)

API key not valid (HTTP 400 Bad Request)

Authentication failed (HTTP 401 / 403)

Connection refused / timeout

Model not found (HTTP 404)

Token / context window limit exceeded

Malformed JSON / screening errors

API key not found / empty key

SSL / TLS certificate error

Slow inference / timeouts

Out of memory (OOM)

Still stuck?

Local & Free AI

Google AI Studio (Gemini)

Mistral AI (La Plateforme)

Z.AI (GLM)

OpenAI

Anthropic

Use a free key in Bango

Why Run LLMs Locally?

Hardware Requirements

Ollama Setup (Recommended)

LM Studio Setup

llama.cpp Setup (Advanced)

Recommended Models

Configuration Tips for Bango

Further Reading

Feature Reference

How Bango Works with AI

Bango RAG vs. Dumping Documents into Chat Windows

Benefits & Drawbacks of Bango's Approach

Key Features

Key Features

Screen Controls & Buttons

Key Features

Format Examples

Deduplication Strategies

Screen Controls & Buttons

Tags vs. Labels

Screen Controls & Buttons