RAG Console

Minimal Flask RAG app for uploading a document, indexing it locally, and asking questions with citations.

Features

• Upload and index a single source (PDF/DOCX/TXT/MD/CSV/JSON/HTML/XML).
• Hybrid retrieval: dense embeddings (Gemini) + BM25, fused with RRF.
• Top matches view plus grounded answers with numbered citations.
• Evidence preview: click a citation or a match card to open the referenced chunk.
• Ingestion progress UI (upload, extract, chunk, embed, index).
• Optional: use Qdrant for the dense vector index.

Architecture

flowchart TD
  UI[Browser UI] -->|upload/query| API[Flask app]
  API -->|/api/upload| X[Extract text]
  X --> C[Chunker]
  C --> E[Gemini embed]
  E --> V[(Dense vector index: Qdrant or in-memory)]
  C --> B[(In-memory BM25)]
  API -->|/api/retrieve| R[Hybrid retrieve + RRF fuse]
  V --> R
  B --> R
  API -->|/api/answer| L["Gemini chat (grounded answer)"]
  R --> L
  L --> UI

Loading

Note: architecture.md may be out of date; this README reflects the current code paths.

Configuration

Copy .env.example to .env. Key settings:

• GOOGLE_API_KEY (required)
• MAX_UPLOAD_MB (default 250)
• CHUNK_MAX_TOKENS (default 512)
• CHUNK_OVERLAP_TOKENS (default 32)
• GEMINI_EMBED_MODEL (default text-embedding-004)
• GEMINI_CHAT_MODEL (default gemini-2.5-flash)
• MAX_EMBED_REQUESTS_PER_MINUTE (default 0 = disabled; limits embedding API requests/minute)
• FLASK_HOST (default 0.0.0.0)
• FLASK_PORT (default 5000)
• FLASK_DEBUG (default 0)
• QDRANT_URL (optional, e.g. http://localhost:6333)
• QDRANT_COLLECTION (optional, default rag_chunks)
• QDRANT_API_KEY (optional)
• QDRANT_RECREATE_COLLECTION (optional, default 1)

Query tuning defaults live in webapp/config.py (QuerySettings).

Pipeline

1. Upload + extract text (PDF/DOCX/TXT/etc).
2. Chunk + embed (Gemini), build a dense index (Qdrant or in-memory) + in-memory BM25.
3. Query uses hybrid retrieval + RRF fusion, then the LLM answers with chunk citations (shown in the UI).

Important: the app still keeps chunk text/ordering in process memory for the active upload. Qdrant currently replaces only the dense similarity search component.

Run

1. Create a virtualenv and install deps:
   • python3 -m venv .venv
   • source .venv/bin/activate
   • pip install -r requirements.txt
2. Set your Gemini key:
   • cp .env.example .env
   • edit .env and set GOOGLE_API_KEY
3. (Optional) Start Qdrant locally:
   • docker run -d --name rag-qdrant -p 6333:6333 -v "$(pwd)/qdrant_storage:/qdrant/storage" qdrant/qdrant
   • set QDRANT_URL=http://localhost:6333 in .env
4. Start the server:
   • python3 app.py
5. Open:
   • http://localhost:5000