CodeMind — AI-Powered Code Knowledge Base

Upload your codebase, ask questions in plain English, and get cited answers.
Built with Endee vector database at its core.

---

✨ Features

#	Feature	Description	Endee Usage
1	Ingest Codebase	Upload ZIP or individual code files. System chunks, embeds, and indexes everything.	index.upsert() — store code chunk vectors with metadata
2	RAG Chat	Ask questions about your code in plain English, get cited answers.	index.query() — retrieve relevant chunks for context
3	Semantic Search	Describe what you're looking for, find relevant files by meaning.	index.query() — similarity search across all chunks
4	Recommendations	Select a file, see similar files from the codebase.	index.query() — mean-vector similarity + filtering
5	Agentic Q&A	Complex questions: agent decomposes → multi-search → synthesize.	index.query() × N — multiple searches per sub-question

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        FRONTEND                                 │
│          Next.js 15 (App Router) + Tailwind + shadcn/ui         │
│                                                                 │
│  ┌───────────────────────────┐  ┌────────────────────────────┐  │
│  │     Public Routes         │  │     Protected Routes       │  │
│  │  / (Landing Page)         │  │  /dashboard                │  │
│  │  /login (Auth Forms)      │  │  Requires valid JWT token  │  │
│  └─────────────┬─────────────┘  └─────────────┬──────────────┘  │
│                │ AuthContext (State & Token)  │                 │
│                └──────────────────────────────┘                 │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────────────┐   │
│  │ Upload Panel │  │   Tab Bar    │  │   File Card          │   │
│  │ • Drag & Drop│  │ Ask|Search|  │  │ • Recommendations    │   │
│  │ • File List  │  │ Agent        │  │ • Similar files      │   │
│  └──────┬───────┘  └──────┬───────┘  └──────────┬───────────┘   │
│         │                 │                      │              │
│         └─────────┬───────┴──────────────────────┘              │
│                   │  Next.js API Routes (proxy w/ Auth)         │
└───────────────────┼─────────────────────────────────────────────┘
                    │ HTTP / SSE (Bearer Token)
┌───────────────────┼─────────────────────────────────────────────┐
│                   │          BACKEND                            │
│                   │     FastAPI (Python)                        │
│                   │                                             │
│  ┌────────────────▼──────────────────────────────────────────┐  │
│  │ main.py — Routes                                          │  │
│  │  POST /auth/register POST /auth/login  GET /auth/me       │  │
│  │  ──────────────────────────────────────────────────────── │  │
│  │  POST /ingest  POST /ask  POST /agent                     │  │
│  │  GET  /search  GET  /recommend  GET  /files               │  │
│  └──┬────────────────┬────────────────┬──────────────────┬───┘  │
│     │                │                │                  │      │
│  ┌──▼──────┐  ┌──────▼──────┐  ┌─────▼──────┐      ┌─────▼───┐  │
│  │ingestion│  │   rag.py    │  │  agent.py  │      │ auth.py │  │
│  │  .py    │  │ Embed→Search│  │ Decompose→ │      │ JWT Gen │  │
│  │ Chunk→  │  │  →Prompt→   │  │ MultiSearch│      │ Bcrypt  │  │
│  │ Embed→  │  │  Ollama     │  │ →Synthesize│      │ Verify  │  │
│  │ Store   │  └──────┬──────┘  └─────┬──────┘      └─────┬───┘  │
│  └──┬──────┘         │               │                   │      │
│     │                │               │      ┌────────────▼────┐ │
│  ┌──▼────────────────▼───────────────▼──┐   │     MongoDB     │ │
│  │   endee_client.py (SDK Wrapper)      │   │ • User profiles │ │
│  │   (Filters data by user_id)          │   │ • Hashed PWs    │ │
│  └───────────────────┬──────────────────┘   └─────────────────┘ │
│                      │                                          │
└──────────────────────┼──────────────────────────────────────────┘
                       │ HTTP
             ┌─────────▼─────────┐      ┌──────────────────┐
             │   Endee Vector DB │      │  Ollama (local)  │
             │   localhost:8080  │      │  localhost:11434 │
             │   384-dim cosine  │      │  codellama/llama3│
             └───────────────────┘      └──────────────────┘

---

🛠️ Tech Stack

Layer	Technology	Purpose
Vector DB	Endee (Docker)	Store & search code embeddings
Embeddings	all-MiniLM-L6-v2 (384-dim)	Convert code to vectors
LLM	Ollama (codellama / llama3)	Generate answers from context
Backend	FastAPI (Python 3.11+)	API server with SSE streaming
Frontend	Next.js 15 + Tailwind + shadcn/ui	Dark-themed developer UI

---

🚀 Setup

Prerequisites

• Docker — for Endee vector DB
• Python 3.11+ — for backend
• Node.js 18+ — for frontend
• Ollama — for local LLM

Environment Variables (.env)

Before starting, create a .env file in the backend/ directory:

# MongoDB & Auth
MONGO_URL=mongodb://localhost:27017
MONGO_DB=codemind
JWT_SECRET=your-super-secret-key-change-me
JWT_ALGORITHM=HS256
JWT_EXPIRY_HOURS=72

# Endee Vector DB
ENDEE_HOST=http://localhost:8080
ENDEE_AUTH_TOKEN=
ENDEE_INDEX_NAME=codemind
ENDEE_DIM=384

# LLM & Embeddings
OLLAMA_HOST=http://localhost:11434
OLLAMA_MODEL=codellama
EMBEDDING_MODEL=all-MiniLM-L6-v2

1. Start Endee & MongoDB

# Start Endee Server
docker run -p 8080:8080 -v endee-data:/data endeeio/endee-server:latest

# Start MongoDB (if not installed locally)
docker run -p 27017:27017 -d mongo

2. Start Ollama

ollama pull codellama    # or: ollama pull llama3
ollama serve             # if not already running

3. Backend

cd backend
python -m venv .venv

# On Mac/Linux:
source .venv/bin/activate
# On Windows:
# .\venv\Scripts\activate

pip install -r requirements.txt

# Verify Endee connection
python test_endee.py

# Start API server
uvicorn main:app --host 0.0.0.0 --port 8000 --reload

4. Frontend

cd frontend
pnpm install
pnpm run dev

Open http://localhost:3000

Troubleshooting

• Endee not reachable (ConnectionRefusedError): Ensure Docker is running and the Endee container is started on port 8080.
• Ollama model not found: Run ollama pull codellama to download the model before starting the server. If using a different model, update OLLAMA_MODEL in your .env.
• MongoDB connection failed: Ensure MongoDB is running locally on port 27017, or update MONGO_URL to point to a cloud cluster like MongoDB Atlas.
• Port conflicts: If ports 3000 (frontend), 8000 (backend), or 8080 (Endee) are in use, stop conflicting services or update your environment variables and API proxy targets accordingly.

---

📁 Project Structure

codemind/
├── backend/
│   ├── main.py              # FastAPI routes
│   ├── config.py            # All configuration
│   ├── endee_client.py      # Endee SDK wrapper
│   ├── ingestion.py         # File parsing → chunking → embedding → storing
│   ├── rag.py               # RAG pipeline (search → prompt → stream)
│   ├── agent.py             # Agentic pipeline (decompose → multi-search → synthesize)
│   ├── test_endee.py        # End-to-end Endee validation
│   └── requirements.txt
└── frontend/
    ├── app/
    │   ├── page.tsx          # Main two-panel layout
    │   ├── layout.tsx        # Root layout with dark theme
    │   └── api/              # 6 proxy routes → backend
    └── components/
        ├── UploadPanel.tsx   # File upload + indexed files list
        ├── ChatPanel.tsx     # RAG chat with streaming
        ├── SearchPanel.tsx   # Semantic search results
        ├── AgentPanel.tsx    # Agentic Q&A with live steps
        └── FileCard.tsx      # File recommendations

---

🔍 How Endee Powers Every Feature

Ingestion

Each code file is split into 60-line chunks with 10-line overlap, embedded using all-MiniLM-L6-v2, and stored via index.upsert() with metadata {file_path, language, chunk_index, text}.

RAG Chat

User question → model.encode() → index.query(top_k=6) → retrieved chunks become LLM context → Ollama generates cited answer.

Semantic Search

Query → embed → index.query(top_k=8) → return ranked chunks with similarity scores. No LLM involved.

Recommendations

Select a file → retrieve all its chunks → compute mean embedding → index.query() → filter out same file → return top-4 similar files.

Agentic Q&A

Complex question → Ollama decomposes into 3 sub-questions → index.query() for each sub-question → deduplicate chunks → Ollama synthesizes comprehensive answer with citations.

---

📝 API Reference

(Data routes require a valid JWT token in the Authorization: Bearer <token> header)

Method	Endpoint	Description	Auth
POST	/auth/register	Create user account	No
POST	/auth/login	Login, get JWT	No
GET	/auth/me	Get current user info	Yes
POST	/ingest	Upload code file or ZIP	Yes
POST	/ask	RAG Q&A (SSE stream)	Yes
POST	/agent	Agentic Q&A (SSE stream)	Yes
GET	/search?q=...&top_k=8	Semantic search	Yes
GET	/recommend?file_path=...&top_k=4	File recommendations	Yes
GET	/files	List all indexed files	Yes
GET	/health	Health check	No

All responses follow: {"success": true, "data": {...}, "error": null}

---

Demonstrating Semantic Search, RAG, Recommendations, and Agentic AI, all powered by the Endee vector database.