Jyotish AI (Vedic-RAG)

An Agentic RAG Streamlit application that grounds Vedic astrology responses in two sources:

• Accurate astronomical data from FreeAstrologyAPI
• Classical interpretive rules from Brihat Parashara Hora Shastra (BPHS) via Pinecone

The app collects DOB, TOB, and City, fetches the correct divisional chart (D1/D9/D10), retrieves relevant BPHS passages, and synthesizes the answer. All SVG charts are rendered directly in the chat.

Architecture

• Orchestration: LangChain Agent (ReAct) + Tools
• Frontend: Streamlit main.py
• LLM Factory: OpenAI / Google Gemini / AWS Bedrock (via env LLM_PROVIDER)
• Embeddings: OpenAI / Gemini / Bedrock via src/embedding_factory.py (env EMBEDDING_PROVIDER)
• Vector DB: Pinecone (Serverless)
• Cache & History: MongoDB Atlas (api_cache, chat_history) using langchain-mongodb
• Prompt Management: LangSmith Prompts (pushed via scripts/setup_prompts.py, loaded via Client in src/agent.py)
• Observability: LangSmith (Tracing enabled)

See detailed design in docs/PRD.md and docs/EDD.md.

Project Layout

vedic-rag/
├── .env.example
├── Dockerfile
├── requirements.txt
├── main.py
├── scripts/
│   ├── ingest.py                           # Ingest BPHS PDF into Pinecone
│   └── setup_prompts.py                    # Push system prompt to LangChain Hub
├── src/
│   ├── config.py                           # Env + config
│   ├── llm_factory.py                      # Factory Method for LLMs
│   ├── utils.py                            # Geocoding + timezone offset
│   ├── vector_store.py                     # Pinecone retriever helper
│   ├── embedding_factory.py                # Embedding provider selection (OpenAI/Gemini)
│   ├── tools.py                            # D1/D9/D10 tools + MongoDB caching + BPHS search
│   └── agent.py                            # AgentExecutor with tools + chat history
├── data/
│   └── brihat-parashara-hora-shastra-english-v.pdf   # Source PDF (example path)
└── docs/
    ├── PRD.md
    └── EDD.md

Setup

1) Python Environment

python -m venv .venv
source .venv/bin/activate  # macOS/Linux
pip install -r requirements.txt

2) Environment Variables

Copy .env.example → .env and fill values:

• LLM_PROVIDER: openai | google_genai | bedrock
• OPENAI_API_KEY / GOOGLE_API_KEY / AWS creds
• PINECONE_API_KEY, PINECONE_INDEX_NAME
• EMBEDDING_PROVIDER: openai | gemini | bedrock
• MONGO_URI, MONGO_DB_NAME, MONGO_API_CACHE_COLLECTION, MONGO_CHAT_HISTORY_COLLECTION
• FREE_ASTROLOGY_API_KEY
• LANGCHAIN_API_KEY (LangSmith tracing)

3) Ingest BPHS into Pinecone

Ensure your BPHS PDF exists at data/BPHS.pdf (or update the path in scripts/ingest.py).

python scripts/ingest.py

This creates/uses the Pinecone serverless index and loads embedded chunks.

Embedding model selection is controlled by EMBEDDING_PROVIDER:

• openai → text-embedding-3-small (dimension 1536)
• gemini → models/text-embedding-004 (dimension 768)
• bedrock → amazon.titan-embed-text-v2:0 (dimension 1024; uses AWS_REGION_NAME, optional AWS_PROFILE)

On free tiers, ingestion uses batching with short sleeps to avoid rate limits.

4) Run the App

streamlit run main.py

Enter Email (Session ID), DOB, TOB, and City. Ask questions about career (D10), marriage (D9), or health (D1).

5) VS Code Debugging

Use [ .vscode/launch.json ] to debug via the Python streamlit module. Set breakpoints in src/agent.py and tools.

LLM Factory

Select provider via LLM_PROVIDER in .env:

• openai → GPT-4o
• google_genai → Gemini 1.5 Pro
• bedrock → Claude 3 Sonnet (AWS Bedrock)

Tools & Caching

Tools in src/tools.py:

• get_d10_chart(dob, tob, city) – career (D10)
• get_d9_chart(dob, tob, city) – marriage (D9)
• get_d1_chart(dob, tob, city) – general health (D1)
• get_d2_chart(dob, tob, city) – wealth/family (D2)
• get_d7_chart(dob, tob, city) – progeny/children (D7)
• get_d24_chart(dob, tob, city) – education/knowledge (D24)
• get_specific_varga_chart(dob, tob, city, chart_code) – advanced charts by code
• search_bphs(query) – search BPHS via Pinecone

MongoDB caching keys: dob+tob+lat+lon+chart_type. Checks api_cache collection before calling FreeAstrologyAPI.

Chat history is stored per-session (email) using MongoDBChatMessageHistory from langchain-mongodb.

Example session configuration is set in main.py and passed into the agent executor to isolate histories.

Docker

Build and run with env:

docker build -t jyotish-ai:latest .
docker run --rm -p 8501:8501 --env-file .env jyotish-ai:latest

Streamlit will be available at http://localhost:8501.

Notes

• SVG charts must be rendered with st.markdown(svg, unsafe_allow_html=True).
• Prompts are managed via LangSmith; push with scripts/setup_prompts.py and load in src/agent.py.
• For production, consider Auth0 for OAuth and deployment on Streamlit Cloud or AWS.

MongoDB Atlas Setup (POC)

1. Create a free M0 cluster in MongoDB Atlas.
2. Add a database user with read/write access.
3. Whitelist your IP (or use 0.0.0.0/0 for dev only).
4. Copy the connection string and set MONGO_URI in .env.
5. Use MONGO_DB_NAME=jyotish_ai_cache and collections api_cache, chat_history.

Chat histories are partitioned by your session identifier (email) for easy retrieval.