AI Document Adjudicator Engine An advanced Retrieval-Augmented Generation (RAG) system designed to analyze complex legal and policy documents with high accuracy and speed. This project provides a robust API for developers to integrate intelligent document analysis into their applications.
Key Features High-Speed Indexing: Utilizes FAISS (Facebook AI Similarity Search) for blazingly fast in-memory vector indexing and retrieval.
State-of-the-Art Accuracy: Implements a retrieve-and-rerank architecture with a Cross-Encoder model to ensure the most semantically relevant context is used for generating answers.
Intelligent & Transparent Reasoning: Powered by Gemini 1.5 Pro, the system can handle vague queries, make and state clear assumptions, and provide detailed, auditable justifications for its decisions.
Structured JSON Output: Returns a predictable, multi-layered JSON response that is ideal for downstream automation and integration into other systems.
Robust API: Built with FastAPI, providing an interactive and easy-to-use interface for uploading documents and submitting queries.
Workflow The system follows a sophisticated RAG pipeline designed to maximize both speed and accuracy.
flowchart TD
subgraph "User Interaction & Document Indexing"
A["📤
1. Upload Document
(.pdf, .docx, .eml)"] --> B["✂
2. Document Chunking
(Semantic Splitting)"]
B --> C["🤖
3. Text Embedding
(Gemini Model)"]
C --> D["🗄
4. Vector Storage
(Index in FAISS)"]
end
subgraph "Live Query Processing"
E["⌨<br><b>5. User Query Input</b>"] --> F["🤖<br><b>6. Query Embedding</b>"]
D -- Provides Stored Context --> G["🔍<br><b>7. Retrieve & Re-Rank</b><br>(FAISS Search + Cross-Encoder)"]
F -- Provides Query Vector --> G
G --> H["📝<br><b>8. Prompt Construction</b><br>(Combine Context + Query)"]
H --> I["🧠<br><b>9. LLM Synthesis</b><br>(Gemini 1.5 Pro)"]
I --> J["📨<br><b>10. API Response</b><br>(Structured JSON)"]
end
classDef indexing fill:#e6f3ff,stroke:#007bff,stroke-width:2px,color:#000
classDef inference fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px,color:#000
classDef io fill:#fff0e6,stroke:#fd7e14,stroke-width:2px,color:#000
classDef merge fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px,color:#000
class A,E io
class B,C,D indexing
class F,H,I inference
class G merge
class J io
Setup and Installation
- Prerequisites Python 3.8 or higher
A Google Gemini API Key
- Installation Clone the repository:
git clone cd
Set up your API Key: You need to set your Gemini API key as an environment variable.
Windows (Command Prompt):
set GEMINI_API_KEY="YOUR_API_KEY_HERE"
Windows (PowerShell):
$env:GEMINI_API_KEY="YOUR_API_KEY_HERE"
macOS / Linux:
export GEMINI_API_KEY="YOUR_API_KEY_HERE"
Install dependencies:
pip install -r requirements.txt
How to Run the API Start the API Server: Run the following command in your terminal from the project's root directory:
uvicorn api:app --reload
The server will start, typically on http://127.0.0.1:8000.
Access the API Documentation: Open your web browser and navigate to http://127.0.0.1:8000/docs. You will see the interactive FastAPI documentation (Swagger UI), which you can use to test the endpoints directly.
API Workflow Example (using curl) Upload a Document: (Replace path/to/your/policy.pdf with the actual file path)
curl -X POST -F "file=@path/to/your/policy.pdf" http://127.0.0.1:8000/documents/
This will return a JSON response with a unique document_id.
Query the Document: (Replace {document_id} with the ID from the previous step)
curl -X POST -H "Content-Type: application/json"
-d '{"query": "46M, knee surgery, 3-month policy"}'
http://127.0.0.1:8000/documents/{document_id}/query/
This will return the final, structured JSON analysis.