🧠 EDIS — Enterprise Document Intelligence System

A production-grade multi-agent RAG system that ingests any document format, answers questions with citations, and evaluates its own response quality using RAGAS.

Built with LangGraph, LlamaIndex, Qdrant, and a BYOK (Bring Your Own Key) LLM layer supporting Ollama, OpenAI, and Anthropic.

---

Architecture

User (Streamlit UI)
        ↓
   FastAPI Backend
        ↓
   LangGraph Orchestrator
    ┌───────────────────────────────────┐
    │  Ingestion Agent                  │
    │  → PDF/DOCX/CSV/TXT/URL parser   │
    │  → Semantic + fixed chunking     │
    │  → Embedding → Qdrant index      │
    ├───────────────────────────────────┤
    │  Retrieval Agent                  │
    │  → Semantic search (Qdrant)      │
    │  → Cross-encoder reranking       │
    ├───────────────────────────────────┤
    │  Synthesis Agent                  │
    │  → Grounded answer generation    │
    │  → Inline [Context N] citations  │
    ├───────────────────────────────────┤
    │  Evaluation Agent                 │
    │  → RAGAS: Faithfulness           │
    │  → RAGAS: Answer Relevancy       │
    │  → RAGAS: Context Precision      │
    └───────────────────────────────────┘
        ↓
   BYOK Layer (litellm)
   Ollama | OpenAI | Anthropic

---

Features

• Multi-format ingestion — PDF, DOCX, CSV, TXT, Markdown, and web URLs
• Two chunking strategies — semantic (paragraph-aware) and fixed (token-based)
• Cross-encoder reranking — improves retrieval precision significantly over naive top-k
• Grounded answers with citations — every answer references specific context chunks
• RAGAS evaluation — faithfulness, answer relevancy, context precision scored per query
• BYOK LLM routing — litellm routes to Ollama (default), OpenAI, or Anthropic
• Configurable embeddings — nomic-embed-text (local) or text-embedding-3-small (OpenAI)
• LangGraph orchestration — clean state machine, each agent is a discrete node
• REST API — full FastAPI backend, usable independently of the UI

---

Tech Stack

Layer	Tool
Orchestration	LangGraph
RAG framework	LlamaIndex
Vector DB	Qdrant (Docker)
Reranker	cross-encoder/ms-marco-MiniLM-L-6-v2
LLM routing	litellm
Evaluation	RAGAS
Backend	FastAPI
Frontend	Streamlit
Document parsing	PyMuPDF, python-docx, pandas, trafilatura

---

Quickstart

1. Clone and install

git clone https://github.com/akshaygande/edis.git
cd edis
pip install -r requirements.txt

2. Configure environment

cp .env.example .env
# Edit .env with your preferred LLM provider and keys

3. Start Qdrant

docker compose up -d

4. Pull embedding model (Ollama default)

ollama pull nomic-embed-text
ollama pull llama3.2

5. Start the backend

cd backend
uvicorn main:app --reload --port 8000

6. Start the frontend

cd frontend
streamlit run app.py

Open http://localhost:8501

---

Configuration

All config is driven by .env. Key variables:

Variable	Default	Description
LLM_PROVIDER	ollama	ollama / openai / anthropic
EMBEDDING_PROVIDER	ollama	ollama / openai
CHUNKING_STRATEGY	semantic	semantic / fixed
CHUNK_SIZE	512	Words per chunk
TOP_K	10	Chunks retrieved before reranking
RERANK_TOP_N	4	Chunks kept after reranking

---

API Reference

Endpoint	Method	Description
/health	GET	Health check
/ingest	POST	Upload and ingest a file
/ingest/url	POST	Ingest a web URL
/query	POST	Ask a question
/collection/info	GET	Qdrant collection stats
/collection	DELETE	Wipe all indexed data

Example query

curl -X POST http://localhost:8000/query \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What are the main risk factors?",
    "skip_evaluation": false
  }'

Example response

{
  "query": "What are the main risk factors?",
  "answer": "Based on the documents, the main risk factors include... [Context 1][Context 3]",
  "citations": [
    {
      "index": 1,
      "source": "annual_report.pdf",
      "page_num": 12,
      "score": 0.91,
      "text_preview": "The company faces significant exposure to..."
    }
  ],
  "retrieval_stats": { "total_retrieved": 10, "total_reranked": 4, "duration_seconds": 0.8 },
  "synthesis_stats": { "model_used": "ollama/llama3.2", "provider": "ollama", "duration_seconds": 2.1 },
  "evaluation": {
    "faithfulness": 0.92,
    "answer_relevancy": 0.88,
    "context_precision": 0.85
  }
}

---

Project Structure

edis/
├── backend/
│   ├── main.py                  # FastAPI entry point
│   ├── config.py                # Central config + BYOK routing
│   ├── agents/
│   │   ├── ingestion_agent.py   # Parse → chunk → embed → index
│   │   ├── retrieval_agent.py   # Semantic search + reranking
│   │   ├── synthesis_agent.py   # LLM answer generation
│   │   └── evaluation_agent.py  # RAGAS evaluation
│   ├── graph/
│   │   └── orchestrator.py      # LangGraph state graph
│   ├── parsers/
│   │   ├── pdf_parser.py
│   │   ├── docx_parser.py
│   │   ├── csv_parser.py
│   │   ├── txt_parser.py
│   │   ├── url_parser.py
│   │   └── __init__.py          # Parser router
│   ├── rag/
│   │   ├── embedder.py
│   │   ├── indexer.py
│   │   ├── retriever.py
│   │   └── reranker.py
│   └── evals/
│       └── ragas_eval.py
├── frontend/
│   └── app.py                   # Streamlit UI
├── docker-compose.yml
├── requirements.txt
├── .env.example
└── README.md

---

Why EDIS stands out

Most RAG portfolios are: load PDF → top-k search → dump into prompt → answer. EDIS goes further:

1. Reranking — cross-encoder reranking on top of semantic search cuts hallucination significantly
2. Multi-agent architecture — each stage is a discrete, testable LangGraph node
3. Self-evaluation — RAGAS scores tell you how good the answer is, not just what it is
4. BYOK — works locally with Ollama, no API key required
5. Production patterns — FastAPI backend, Docker Qdrant, config-driven, no hardcoded secrets

---

License

MIT