Retrieve first, generate second
Lexical retrieval narrows the context to relevant excerpts. The LLM receives only the question and those excerpts; its system prompt says to treat document content as untrusted and not follow instructions inside it.
KNOWLEDGE ASSISTANT · AI / RAG
How documents
become answers.
The assistant extracts text, retrieves relevant passages, and sends only those passages with the question to a generative model. The answer includes source filenames, page numbers, and citations.
QUICK DEMO · 22 SECONDS
See the workflow.
The video shows a question about the built-in sample and an answer with a source. The app also supports uploads after account creation; free demo data may reset when the service restarts.
Download the short video ↗01 / DATA FLOW
From file to answer.
- 01 · WEB CLIENTReact
The demo works without signing in on a built-in sample; PDF, DOCX, and TXT uploads go to the signed-in account library.
- 02 · EXTRACTIONFastAPI + parsers
Validates type and size, extracts text, and preserves page numbers. The temporary file is removed after processing.
- 03 · PASSAGESChunking + SQLite
Normalizes text, creates 1,100-character chunks with 160-character overlap, and stores text with source metadata.
- 04 · RETRIEVALLexical search
Compares question and passage terms, ranks matches with cosine similarity, and selects up to three sources.
- 05 · GENERATIONOpenRouter · free route
The server sends the question and retrieved excerpts to the model. The API key stays in Render environment variables and never reaches the browser.
- 06 · RESULTAnswer + sources
The UI displays the answer, cited passages, and recent questions; chat history is stored in SQLite.
Data boundary: the original uploaded file is not sent to OpenRouter. The provider receives the question and relevant extracted passages — up to about 6,000 context characters per request.
02 / DESIGN DECISIONS
What happens inside.
Answers you can verify
Each API source includes a filename, page, excerpt, and relevance score. The model is asked to cite sources so readers can check claims against the document.
The key stays server-side
React calls the app's own API. FastAPI reads OPENROUTER_API_KEY from the environment, never from the browser or public repository.
Synchronous path for a small demo
Indexing runs during the upload request. This keeps the demo simple; large files and higher traffic need a task queue and resource limits.
03 / DEPLOYMENT PROFILES
Hosted demo and full Compose profile.
| Layer | Public Render demo | Full Compose profile |
|---|---|---|
| UI / API | React · FastAPI | React · FastAPI |
| Data | SQLite · container filesystem | PostgreSQL · upload volume |
| Search / generation | Lexical search · OpenRouter Free route | Qdrant vectors · Ollama embeddings and LLM |
| Background services | Requests run in a single web service | Compose includes Redis and a Celery worker; indexing can be moved into background tasks |
| Persistence | Data may reset after a restart or idle period | Local Docker volumes preserve data across restarts |
The free profile lets visitors try the app without installation. The full profile is for exploring vector search and a local model; it is a development option, not a separate public production service.
04 / KNOWN GAPS
Before using real customer data.
Persistent database: the app supports PostgreSQL, but the current successful demo release uses ephemeral SQLite on Render.
Files and backups: source uploads are deleted after indexing; object storage and a tested database restore process are still needed.
Already implemented: accounts, per-account document and history isolation, and owner-only document deletion.
Next improvements: email verification, password recovery, queued indexing, and retrieval/citation quality evaluation.