AI-Powered Document Parsing and Reporting Tool

Project Type: Web app / AI tool Date: 1st October, 2025 Stack / Tools Used:

Python, Streamlit (frontend)

Unstructured (for document parsing)

LangChain / Google Generative AI (or OpenAI) for embeddings & LLM

Vector store or in-memory retrieval

Hosting / deployment (specify e.g. streamlit cloud)

When working with multiple documents (PDFs, Word files, PPTs), extracting insights, summarizing, or generating structured reports is tedious and manual. We wanted a tool where users can upload documents, interact with them in natural language, and automatically generate executive reports or answer ad hoc questions — all in a conversational ChatGPT style.

Key pain points:

Manual skimming and copying across documents

Difficulty navigating long documents

Inconsistent formats and missing structure

Lack of a tool that combines parsing + LLM reasoning in chat form

To build this, I focused on modular architecture:

The application is a single-page Streamlit app (app.py) that handles file upload, parsing, embeddings, retrieval, and chat all in one flow.

To avoid re-parsing or re-embedding on every user message, I maintain state in st.session_state (document text, chunk embeddings, message history).

For embeddings, I integrated Google Generative AI embeddings (via LangChain) and for conversational response, used ChatGoogleGenerativeAI (Gemini).

Retrieval is done via a simple cosine similarity ranking over embeddings. For production, this can be swapped out for a vector database (FAISS, Chroma, or GCP Vertex).

The prompt encourages the model to return a structured report with named sections, making it easier to read and parse further.

For documents containing tables, I extract table HTML and render previews in the sidebar so users can see extracted structured data.

Users can upload a document and get a report in seconds, vs. manually reading and summarizing in minutes

The system handles multi-page documents with tables, producing coherent executive summaries

Chat interface allows follow-up questions (e.g. “Highlight KPIs from the doc”, “Which section supports claim X”)

Parsing + embedding reuse ensures performance stays reasonable even with multiple questions per document

Average report generation latency = 2.3s over 10 test docs

Challenges: Parsing complex PDF tables

Solution / Mitigation: Used Unstructured’s table metadata and fallback strategies; display previews for verification

Challenges: Prompt length / token limits

Solution / Mitigation: Chunking + retrieval ensures only relevant context is passed to the model

Challenges: API key security & rate limits

Solution / Mitigation: Kept keys in environment, used conditional initialization, cached embeddings

Challenges: Deployment & file limits

Solution / Mitigation: Configured streamlit upload size, used temp storage, cleaned up intermediates

This project demonstrates how to bridge document processing + conversational AI into a user-friendly app. It’s a great showcase of:

integrating open-source parsing (Unstructured)

building RAG pipelines

using chat models for structured output

designing UI/UX for conversational report generation

It’s now part of my portfolio to show how I build end-to-end AI applications from file ingestion to report output.