Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Jatin-Mehra119/PDF-Insight-Beta/llms.txt

Use this file to discover all available pages before exploring further.

PDF Insight Pro is a production-ready document intelligence API that combines Retrieval Augmented Generation (RAG) with LangChain tool-calling agents. Upload any PDF, then query it in natural language — the system retrieves semantically relevant chunks, invokes a Groq-powered LLM, and optionally validates answers against live web results via Tavily.

Quickstart

Run PDF Insight Pro locally and make your first API call in minutes.

Configuration

Set API keys, tune chunk sizes, select models, and configure FAISS parameters.

Architecture Overview

Understand the layered architecture — API routes, services, utilities, and storage.

API Reference

Full reference for all REST endpoints: upload, chat, sessions, and models.

How It Works

PDF Insight Pro processes your documents through a four-stage pipeline:
1

Upload a PDF

POST /upload-pdf — the server extracts text with PyMuPDF, splits it into semantic chunks, generates embeddings with BAAI/bge-large-en-v1.5, and indexes them in a FAISS HNSW index. A session ID is returned.
2

Ask a question

POST /chat — send your session ID and a natural-language query. The agent retrieves the top matching chunks and passes them as context to the Groq LLM.
3

Agent reasoning

A LangChain tool-calling agent decides whether the initial context is sufficient or whether it needs to perform additional vector searches or Tavily web lookups.
4

Receive a grounded answer

The LLM synthesises an answer from document context (and optionally live web results). Chat history is persisted to the session for multi-turn conversations.

Key Features

Agentic RAG

LangChain create_tool_calling_agent drives multi-step retrieval — the agent can call vector_database_search and tavily_search_results_json autonomously.

FAISS HNSW Index

Fast approximate nearest-neighbour search with configurable efConstruction and efSearch parameters for precision/speed trade-offs.

Multiple LLMs via Groq

Switch between Llama 4 Scout 17B, Llama 3.1 8B Instant, Llama 3.3 70B Versatile, and GPT-OSS 120B at upload time.

Web Search Augmentation

Toggle use_search: true in any chat request to let the agent verify document content against live Tavily web results.

Session Persistence

Sessions survive restarts — data is serialised to disk and reconstructed on demand, keeping the FAISS index and chat history intact.

Docker & HF Spaces

First-class Docker support and a GitHub Actions workflow for zero-config deployment to Hugging Face Spaces.

RAG System Performance

The RAG pipeline has been evaluated on the neural-bridge/rag-dataset-12000 benchmark:
MetricValue
Semantic Similarity (mean)0.852
ROUGE-L F1 (mean)0.395
Evaluation pass rate85.3% (64 / 75)
See the test_RAG.ipynb notebook in the repository for the full evaluation methodology and quantile distributions.

Build docs developers (and LLMs) love