AI-Powered Document Query Engine

An advanced domain-agnostic document analysis system built with FastAPI, LangGraph, and Qdrant.
Easily ingest documents, ask multiple natural language questions, and receive structured, verifiable answers — all in one API call.

---

✨ Key Highlights

• 📍 One API to Rule Them All – A single /process endpoint handles document ingestion + batch querying in one go.
• 🌎 Works Across Domains – Insurance, Legal, HR, Finance — the system detects the domain and extracts relevant entities automatically.
• 🔒 Isolated Vector Collections – Each job has its own Qdrant collection (jobId) for data segregation.
• ⚡ Parallel Question Processing – Handles multiple queries simultaneously for high throughput.
• 🧠 Advanced RAG Pipeline:
  1. Retrieval – Get broad, relevant document chunks.
  2. Reranking – Refine with a CrossEncoder for precise context.
  3. Generation – Multi-persona LLM (Analyst + Auditor) produces structured answers with self-critique + confidence score.
• 📂 Multi-Format Support – Works with PDF, DOCX, and EML files.

---

🏗 Architecture

Powered by a LangGraph-based graph workflow:

1. FastAPI Server – Receives API requests and orchestrates the workflow.
2. LangGraph Workflow – A sequence of stateful nodes ensuring debuggable and maintainable execution.
3. Qdrant Vector DB – Stores vectorized document chunks per jobId.
4. LangChain + OpenAI – Handles question analysis, structured answer generation, and self-critique.
5. Sentence Transformers – Reranks retrieved documents for maximum relevance.
6. Document Parsers – PyMuPDF, python-docx, and mailparser extract clean text from various formats.

---

🔍 Workflow Overview

The /process endpoint triggers this sequence:

1. preprocess – Download, parse, and chunk the document.
2. batch_analyze_queries – Identify document domain + generate search queries for each question.
3. load_to_db – Store chunks in a job-specific Qdrant collection.
4. batch_retrieve_docs – Retrieve broad context for all questions.
5. batch_rerank_docs – Use CrossEncoder to create tailored context per question.
6. batch_generate_answers – Generate structured answers + self-critique.

---

📡 API Usage

POST /process

Ingest a document, process questions, and receive structured answers.

Request Body:

{
  "jobId": "string",
  "documents": "string (URL)",
  "questions": ["string"]
}

• jobId – Unique identifier, used as Qdrant collection name.
• documents – Public URL to PDF/DOCX/EML file.
• questions – List of natural language questions.

Response:

{
  "answers": [
    {
      "decision": "string",
      "details": {},
      "justification": "string",
      "clauses": ["string"]
    }
  ]
}

• decision – Short, direct outcome (e.g., "Approved").
• details – Extracted facts (e.g., amount, waiting period).
• justification – Step-by-step reasoning based on document.
• clauses – Supporting excerpts or clause IDs.

---

⚙️ Setup

Prerequisites

• Python 3.9+
• Docker (for Qdrant)
• OpenAI API key

1. Clone the Repository

git clone <your-repo-url>
cd <your-repo-name>

2. Install Dependencies

pip install -r requirements.txt

3. Configure Environment

Create .env:

OPENAI_API_KEY="sk-..."
QDRANT_URL="http://localhost:6333"
QDRANT_API_KEY=null
EMBEDDING_MODEL="text-embedding-3-small"

4. Run Qdrant (via Docker)

docker run -p 6333:6333 -p 6334:6334 qdrant/qdrant

---

▶️ Run the Application

uvicorn main:app --reload

Access:

• API – http://127.0.0.1:8000
• Docs – http://127.0.0.1:8000/docs

---

💻 Tech Stack

Component	Technology
Backend	FastAPI
Workflow Orchestration	LangGraph
RAG/LLM Framework	LangChain
Vector Database	Qdrant
LLM Provider	OpenAI
Reranker	Sentence Transformers (CrossEncoder)
Document Parsing	PyMuPDF, python-docx, mailparser

---

🏆 Why This Matters

This engine isn't just a Q&A bot — it’s a scalable, auditable, and high-precision document understanding system that adapts to any industry, providing reliable and explainable answers at scale.