Inductify

An AI-powered employee onboarding assistant that answers company policy questions with cited source excerpts. Built with FastAPI, LangChain, ChromaDB, and Next.js.

Measured accuracy: 1.9% (LLM-only) → 83.3% (RAG) on a 54-question held-out eval set.

---

Quick Start (Docker)

Prerequisites: Docker Desktop, an OpenAI API key (or a local Ollama instance).

# 1. Clone the repo
git clone <repo-url>
cd Inductify

# 2. Set your API key
cp .env.example .env
# Edit .env and set OPENAI_API_KEY=sk-...

# 3a. Empty database — upload your own documents via the UI
docker compose up --build

# 3b. Pre-loaded demo mode — 5,500 synthetic policy documents indexed automatically
docker compose --profile demo up --build

Demo mode calls the OpenAI Embeddings API on first run (~$0.05 for the full corpus).
Watch the indexer container logs; once it prints ✓ Done, all policy questions are answerable.
Subsequent docker compose --profile demo up runs reuse the persisted ChromaDB volume — no re-indexing.

Service	URL
Frontend	http://localhost:3000
Backend	http://localhost:8000
API docs	http://localhost:8000/docs

---

Features

• RAG Q&A — retrieves the top-k relevant chunks from ChromaDB, feeds them as context to the LLM, and returns an answer with numbered source citations
• Conversation history — multi-turn sessions tracked server-side; 8 pre-seeded demo conversations shown in the sidebar
• Document upload — drag-and-drop .txt, .md, .pdf, and .xlsx files; backend vectorizes and indexes asynchronously
• ReAct agent — optional tool-calling agent (/agent/ask) for multi-step reasoning
• Dark / light mode — toggle in the chat header
• Ollama fallback — runs fully offline when OPENAI_API_KEY is unset
• Optional reranker — CrossEncoder reranking available as an opt-in (adds ~2.5 GB to the image)

---

Using a Local Model (No OpenAI Key)

Install Ollama, pull a model, then leave OPENAI_API_KEY empty:

ollama pull llama3.2

# .env
OPENAI_API_KEY=
OLLAMA_MODEL=llama3.2
OLLAMA_BASE_URL=http://host.docker.internal:11434

---

Uploading Your Own Documents

1. Open the chat UI at http://localhost:3000
2. Click the paperclip icon → select .txt, .md, .pdf, or .xlsx files
3. The backend vectorizes and stores them; a status badge shows indexing progress
4. Ask questions — answers include the source filename and a relevant excerpt

---

Optional: CrossEncoder Reranker

The reranker re-scores retrieved chunks before generation, improving answer quality at the cost of ~2.5 GB of additional image size (PyTorch + sentence-transformers).

# Enable at build time
INSTALL_RERANKER=true docker compose up --build

# Or in .env:
INSTALL_RERANKER=true
DISABLE_RERANKER=       # leave empty to activate reranking at runtime

By default DISABLE_RERANKER=1 so the backend works without the reranker installed.

---

Project Structure

Inductify/
├── backend/
│   ├── llm/               # LLM factory (OpenAI / Ollama), RAG chain, ReAct agent
│   ├── embedding/         # OpenAI embeddings + optional CrossEncoder reranker
│   ├── database/          # ChromaDB manager, file loaders (txt/md/pdf/xlsx)
│   ├── scripts/           # index_demo_data.py — bulk indexer for demo profile
│   ├── auth.py            # API key auth + rate limiting
│   └── main.py            # FastAPI endpoints
├── frontend/              # Next.js 15 + HeroUI chat interface
│   └── components/        # Sidebar with history, message bubbles, file upload
├── eval/                  # Offline evaluation pipeline (not used in production)
│   ├── rag_eval.py        # 3-condition ablation: LLM-only / RAG / RAG+rerank
│   ├── build_policy_db.py # Builds policy-only ChromaDB for eval
│   └── rag_eval_builder/  # Synthetic corpus + Q&A pair generator
├── data/                  # Eval datasets, synthetic docs, results
├── docs/                  # Architecture notes
├── docker-compose.yml
└── .env.example

---

API Endpoints

Method	Endpoint	Description
POST	/ask	RAG Q&A with source citations
POST	/agent/ask	ReAct agent (tool-calling)
POST	/upload	Upload documents
POST	/index	Trigger vectorization of uploads
GET	/index/status/{job_id}	Indexing progress
DELETE	/session/{id}	Clear conversation history
GET	/health	Health check

---

Configuration

All options are set via environment variables. See .env.example for the full list.

Variable	Default	Description
OPENAI_API_KEY	—	OpenAI key; leave empty or disabled to use Ollama
LLM_MODEL	gpt-4.1-mini	OpenAI model name
OLLAMA_MODEL	llama3.2	Ollama model (used when no OpenAI key)
OLLAMA_BASE_URL	http://localhost:11434	Ollama server address
INSTALL_RERANKER	false	Set true at build time to install CrossEncoder
DISABLE_RERANKER	1	Set empty to enable reranker at runtime (requires build)
API_KEY	disabled	Backend API key (disabled = no auth)
ALLOWED_ORIGINS	*	Comma-separated CORS origins
POLICY_ONLY	1	Demo mode: 1 = policy docs only, 0 = full corpus

---

Tech Stack

• Backend: Python · FastAPI · LangChain · LangGraph · ChromaDB · OpenAI / Ollama
• Frontend: Next.js 15 · React · TypeScript · HeroUI · Tailwind CSS
• Infra: Docker Compose (multi-service with health-check dependency chain)

License

MIT