Krishna Saarthi — Backend

Go backend for Krishna Saarthi, a Bhagavad Gita-inspired philosophical guide powered by Sarvam AI.

Stack

• Go — HTTP server (net/http)
• Sarvam AI (sarvam-105b) — LLM for response generation
• Supermemory — Long-term memory per user across sessions
• MongoDB Atlas — Persists users, sessions, and messages
• Clerk — JWT-based authentication
• Pinecone — Vector database for Gita verse retrieval (RAG)
• Ollama (nomic-embed-text) — Local embeddings for RAG queries (768 dimensions)

Project Structure

backend/
├── cmd/
│   ├── server/main.go          # Entry point, route registration
│   └── ingest/main.go          # One-time script to ingest Gita verses into Pinecone
├── constants/                  # System prompt
├── data/
│   ├── verse.json              # 701 Gita verses (Sanskrit + transliteration)
│   ├── translation.json        # English translations (Swami Gambirananda)
│   └── chapters.json           # 18 chapter summaries
├── internal/
│   ├── db/mongo.go             # MongoDB connection
│   ├── handlers/
│   │   ├── chat.go             # POST /chat handler
│   │   └── health.go           # GET /health handler
│   ├── models/                 # Request/response structs
│   ├── services/
│   │   ├── sarvam.go           # Sarvam AI client
│   │   ├── supermemory.go      # Supermemory client
│   │   ├── chat_store.go       # MongoDB read/write logic
│   │   └── rag.go              # Pinecone + Ollama RAG service
│   └── session/                # In-memory session store (legacy)
└── middlewares/
    ├── auth.go                 # Clerk JWT verification
    └── cors_middleware.go      # CORS headers

Request Flow

POST /chat
  → CORS middleware
  → Auth middleware (verify Clerk JWT, extract userID)
  → Search Supermemory for relevant long-term user context
  → Embed user message via Ollama (nomic-embed-text)
  → Query Pinecone for top 3 relevant Gita verses (RAG)
  → Load last 10 messages for session from MongoDB
  → Build prompt: system + memory context + verse context + history
  → Call Sarvam AI
  → Save user + assistant messages to MongoDB
  → Store exchange in Supermemory (async goroutine)
  → Return response

Routes

Method	Path	Auth	Description
GET	/health	No	Health check
POST	/chat	Yes	Send a message
POST	/auth/sync	Yes	Upsert user on first sign-in
GET	/sessions	Yes	List all sessions for a user
GET	/sessions/{id}/messages	Yes	Load messages for a session

Headers (POST /chat)

Header	Required	Description
X-Session-ID	Yes	UUID per browser tab/session
Authorization	Yes	Bearer <clerk_jwt>

Environment Variables

Create a .env file in the project root:

SARVAM_API_KEY=
SUPERMEMORY_API_KEY=
CLERK_SECRET_KEY=sk_test_
MONGO_DB_URI=mongodb+srv://user:password@cluster.mongodb.net/krishna-saarthi
PINECONE_API_KEY=
PINECONE_INDEX_URL=https://bhagavad-gita-xxxx.svc.us-east-1-aws.pinecone.io
OLLAMA_URL=http://localhost:11434

Getting Started

# Install dependencies
go mod tidy

# Start Ollama (required for RAG embeddings)
ollama serve
ollama pull nomic-embed-text

# Run the server
go run cmd/server/main.go

Server starts on :8080.

RAG — Gita Verse Ingestion

The ingestion script embeds all 701 verses + 18 chapter summaries into Pinecone. This is a one-time setup:

# Copy data files to backend/data/
# Then run:
go run cmd/ingest/main.go

This ingests 719 vectors total (701 verses + 18 chapters) using nomic-embed-text (768 dimensions) via Ollama. Each verse is stored with its translation (Swami Gambirananda), transliteration, word meanings, chapter number, and verse reference (e.g. BG 2.47).

At query time, the user's message is embedded with the same model and the top 3 most semantically similar verses are retrieved and injected into the system prompt.

Memory Architecture

Krishna Saarthi uses a three-layer memory system:

Layer	Storage	Scope	Purpose
Short-term	MongoDB	Per session (last 10 messages)	In-conversation context
Long-term	Supermemory	Per user (across all sessions)	User identity, patterns, history
Knowledge	Pinecone	Global (all users)	Verified Gita verses for RAG

MongoDB Collections

Collection	Description
users	Created on first sign-in via /auth/sync
sessions	One per browser session, keyed by sessionId
messages	All chat messages with userId, sessionId, role, content

Session Persistence

• Each browser tab generates a unique X-Session-ID (UUID)
• The active session ID is persisted in localStorage so refreshing restores the last conversation
• Draft input is also persisted in localStorage and restored on refresh
• Past sessions are listed in the sidebar and fully reloadable

Adding a New Route

1. Add the handler in internal/handlers/
2. Register it in cmd/server/main.go
3. Wrap with AuthMiddleware if it requires authentication
4. Add any new request/response structs to internal/models/