FinSight RAG — AI-Powered Investment Document Q&A Engine

Upload financial documents (PDFs, DOCX, TXT) and ask natural language questions.
Every answer is grounded in your documents and includes source citations. Features a new dual-mode engine: Fast responses or High-Accuracy Cohere reranking.

---

Features

• Document Ingestion — PDF, DOCX, TXT via drag-and-drop upload
• Intelligent Chunking — 512-token chunks with 50-token overlap
• FastEmbed Embeddings — BAAI/bge-small-en-v1.5 for semantic search
• Grounded Q&A — Groq (Llama 3.1) answers strictly from your document context
• Fast & Accurate Modes — Choose between instantaneous retrieval ("Fast") or advanced semantic reranking using Cohere's API ("Accurate") for complex queries.
• Source Citations — Every answer shows the filename + page number
• Confidence Scores — High / Medium / Low based on cosine similarity
• Dual Vector DB — Pinecone (cloud) or ChromaDB (local fallback)
• Document Management — List and delete indexed documents

---

Project Structure

finsight-rag/
├── backend/
│   ├── main.py              # FastAPI app (routes, CORS, registry)
│   ├── ingest.py            # Parse → chunk → embed → store pipeline
│   ├── query.py             # RAG query pipeline (embed → search → Cohere rerank → Groq)
│   ├── pinecone_client.py   # Pinecone / ChromaDB dual backend
│   └── requirements.txt     # Python dependencies
├── frontend/
│   ├── public/
│   │   └── favicon.svg
│   ├── src/
│   │   ├── App.jsx          # Root component (layout, header, panels)
│   │   ├── main.jsx         # React entry point
│   │   ├── index.css        # Tailwind + global styles
│   │   ├── services/
│   │   │   └── api.js       # Axios API service layer
│   │   └── components/
│   │       ├── UploadPanel.jsx   # Drag-and-drop upload + document list
│   │       ├── ChatPanel.jsx     # Q&A chat interface (with Fast/Accurate toggle)
│   │       └── AnswerCard.jsx    # Answer + sources + confidence display
│   ├── index.html
│   ├── package.json
│   ├── vite.config.js
│   ├── tailwind.config.js
│   └── postcss.config.js
├── .env.example             # Required environment variables
└── README.md

---

Prerequisites

• Python 3.10+
• Node.js 18+ and npm
• Groq API key — console.groq.com
• Cohere API key — dashboard.cohere.com
• Pinecone API key (optional, or use ChromaDB) — pinecone.io

---

Quick Start

1. Clone and configure

git clone <your-repo-url>
cd finsight-rag

Copy and fill in your environment variables:

cp .env.example backend/.env

Edit backend/.env:

GROQ_API_KEY=gsk-...
COHERE_API_KEY=cohere_...
PINECONE_API_KEY=...
PINECONE_ENVIRONMENT=us-east-1
PINECONE_INDEX_NAME=finsight-rag-384

# Use ChromaDB locally (no Pinecone needed):
# USE_CHROMA=true

---

2. Backend Setup

cd backend

# Create a virtual environment
python -m venv venv

# Activate it
# Windows:
venv\Scripts\activate
# macOS/Linux:
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Start the server
uvicorn main:app --reload

The API will be available at: http://localhost:8000
Interactive docs: http://localhost:8000/docs

---

3. Frontend Setup

Open a new terminal:

cd frontend

# Install dependencies
npm install

# Start the dev server
npm run dev

The app will be available at: http://localhost:5173

---

Using ChromaDB (Local, No Pinecone Required)

If you don't have a Pinecone account, you can run everything locally:

1. In backend/.env, set: USE_CHROMA=true
2. ChromaDB will automatically persist data to ./chroma_store/
3. No additional setup needed!

---

API Reference

POST /upload

Upload and index a document.

Content-Type: multipart/form-data
Body: file=<PDF|DOCX|TXT>

Response: { doc_id, filename, total_chunks, message }

POST /query

Ask a question and get a grounded answer.

{ 
  "question": "What is the net profit?", 
  "filter_doc_id": null,
  "mode": "accurate" 
}

Response: { "answer": "...", "sources": [...], "confidence": "high", "avg_score": 0.87 }

GET /documents

List all indexed documents.

Response: { "documents": [...], "total": 3 }

DELETE /documents/{doc_id}

Remove a document from the index.

Response: { "success": true, "filename": "...", "message": "..." }

---

UI Design

• Theme: Dark navy + gold accent (finance-grade aesthetic)
• Left Panel: Drag-and-drop upload area + indexed document list
• Right Panel: Chat interface with message history and Fast/Accurate modes toggle
• Answer Cards: Answer text + source badges + confidence bar
• Glassmorphism: Frosted glass panels with subtle borders
• Animations: Slide-up entries, shimmer loading, pulse indicators

---

Grounding & Safety

The system prompt enforces strict document-grounded answers:

"Answer ONLY based on the provided context excerpts. If the answer is not present in the context, respond with 'I could not find this in the uploaded documents.'"

Groq (Llama 3.1) is called with temperature=0.1 to minimise hallucination. When "Accurate" mode is used, Cohere's rerank-english-v3.0 ensures only the most semantically relevant top 5 chunks out of the top 20 Pinecone hits are passed into the LLM context.

---

Configuration

Variable	Default	Description
GROQ_API_KEY	—	Required for Llama 3.1 inference
COHERE_API_KEY	—	Required for accurate reranking
PINECONE_API_KEY	—	Required if USE_CHROMA=false
PINECONE_ENVIRONMENT	us-east-1	Pinecone region
PINECONE_INDEX_NAME	finsight-rag-384	Pinecone index name
USE_CHROMA	false	Use local ChromaDB instead

Chunking parameters (in backend/ingest.py):

• CHUNK_SIZE = 512 tokens
• CHUNK_OVERLAP = 50 tokens

---

Troubleshooting

CORS errors — Ensure the backend is running on port 8000 and the Vite proxy is configured.

PINECONE_API_KEY is not set — Either add your key to .env or set USE_CHROMA=true.

Empty answers — Upload documents first before querying. Ensure your API key has sufficient credits.

Slow first query — Pinecone index creation takes ~30 seconds on first run.

---

License

MIT License — see LICENSE for details.