Agentic AI for Insights on PDF Documents (7 Eleven)

This project focuses on developing an agentic AI system capable of autonomously extracting, summarizing, and generating insights from a collection of long, highly technical PDF manuals for 7-Eleven’s Franke commercial coffee machines. The system combines retrieval-augmented generation (RAG) with a multi-agent reasoning pipeline to deliver precise, contextually relevant answers to complex single- and multi-document queries.

Our goal is to demonstrate how an AI system can streamline information discovery, reduce troubleshooting time, and improve real-world operational decision-making across 7-Eleven stores.

👥 Team Members

Name	GitHub Handle	Contribution
Hikmah Mohammed	@hikmahmohammed	Multi-agent pipeline (Planner → Retriever → Synthesizer), multi-document PDF processing, retrieval backbone exploration, project documentation
Aydin Khan	@AydinK7	Implemented the retrieval backbone with regex based PDF cleaning, LLM implementation with prompt engineering, and evaluated retrieval quality using relevant metrics
Emily Wang	@emily-wang22	Developed frontend message bubble and implemented frontend-side API integration
Sarayu Mummidi	@sarayumummidi	Implemented FastAPI backend integrating the agentic workflow to the frontend, Real-time token by token streaming of LLM response, Frontend UI backdrop, and UI design/styling
Harshika Vijayabharath	@HarshikaV	Developed an advanced chat input module supporting controlled text state, keyboard shortcuts, and integrated send-message operations within the chat workflow.
Isabelle Ye	@iaye1334	Developed frontend AI chatbox input module chatpage UI, and integrated the send message operations
Priya Deshmukh	@priyasd-04	Vector database design, pipeline from user query → embedding model → vector database → returning relevant chunks

🎯 Project Highlights

• Built an intelligent RAG-powered agentic assistant for interpreting dense Franke coffee machine manuals

• Designed a multi-agent architecture (Planner → Retriever → Synthesizer) for structured reasoning and grounded answers

• Implemented a high-quality PDF ingestion pipeline using PyMuPDF, OCR fallback, regex cleaning, and metadata tagging

• Developed a vector search system using FAISS for fast, accurate retrieval across thousands of chunks

• Created a polished React UI chat interface with real-time messaging and auto-scroll behavior

• Achieved strong performance: 76% Precision@5, 92% Recall@5, 7.1s end-to-end latency, 4% hallucination rate

👩🏽‍💻 Setup & Installation

1. Clone Repository

git clone https://github.com/sarayumummidi/7-11-agentic-ai

2. Go to the right directory

cd 7-11-agentic-ai

3. Backend Setup

pip install -r requirements.txt

4. Update config.py:

MISTRAL_API_KEY = your_key

5. Run the backend server

uvicorn main:app --reload

6. Frontend Setup

cd frontend
npm install

7. Run the frontend server

npm run dev

The project will be available at http://localhost:5173 to test.

🏗️ Project Overview

According to project documentation, Franke machine manuals are long, dense, and inconsistent—often 70–100+ pages with mixed diagrams, procedures, and troubleshooting steps. Searching manually is slow and error-prone.

Our System Solves This Through:

• Automated ingestion + preprocessing

• Intelligent chunking + metadata tagging

• Vector search with FAISS

• Multi-agent reasoning for grounded, citation-linked answers

• A user-friendly chat frontend that runs in real time

Real-world impact:

• Faster troubleshooting

• Reduced training burden for new store associates

• More consistent & accurate machine maintenance

• Scalable solution for 83,000+ stores

📊 Data Exploration

Datasets

• 5 manuals (A300–A1000)

• 70–100+ pages each

• Mixed text/diagrams → required OCR fallback

Preprocessing Pipeline

• Text extraction via PyMuPDF

• OCR fallback via pytesseract

• Regex cleaning (headers, footers, whitespace)

Chunking

• ~400 tokens per chunk

• 50-token overlap

• Metadata: manual, source_file, source_page, chunk_id

Key EDA Insights

• Manuals vary heavily in structure → chunk size tuning required

• Metadata essential for selecting the correct machine model

• OCR increased text coverage for diagram-heavy pages

Include visuals such as:

• Chunk length histogram

• FAISS embedding similarity heatmap

🧠 Model Development

🔹 Retrieval Backbone

• Robust extraction and cleaning pipeline

• Metadata-enriched chunking

• FAISS vector index (L2 / cosine search)

🔹 Multi-Agent Architecture

• Planner: identifies intent + relevant manuals

• Retriever: pulls FAISS-ranked chunks

• Synthesizer: generates grounded, citation-backed answer via Mistral

🔹 Why Multi-Agent?

• Reduces hallucinations

• Enables cross-document reasoning

• Modular and scalable

📈 Results & Key Findings

Metric	Value
Precision@5	76%
Recall@5	92%
Retrieval Latency	5ms avg, p95 6ms
End-to-End Latency	7.1 seconds
Hallucination Rate	4%

Insights:

• Chunking + FAISS dramatically improves relevance

• Multi-agent output yields clearer, more structured answers

• Latency acceptable for real store usage

Recommended visuals:

• Retrieval accuracy chart

• Latency distribution

• Annotated sample answer

⚠️ Risks, Challenges, & Mitigations

Challenge	Risk	Mitigation
Hallucinations	Incorrect instructions	Citation-enforced synthesis; retrieval-first pipeline
Latency	Multi-agent delay	FAISS tuning, chunk reduction, caching
Frontend/Backend Integration	Render failures	Standardized API schema, error-handling

🚀 Next Steps

Performance

• Bring latency < 10s consistently

• Add retrieval reranking

Features

• Add advanced search & manual-specific filters

• Authentication + stored chat history

Scalability

• Expand ingestion to all manuals

• Cloud deployment pipeline

Model

• Feedback loops for continual improvement

• Explore lightweight fine-tuning

📝 License

This project is for Break Through Tech AI Studio and not licensed for external use.

📄 References

• Final project presentation

• FAISS documentation

• Mistral / OpenAI APIs

• PyMuPDF, Tesseract OCR

🙏 Acknowledgements

We thank:

• Our Challenge Advisor

• Break Through Tech AI program team

• 7-Eleven for providing guidance and context

• Our AI Studio Coach for feedback and support