Chat With Your Files — RAG Document Intelligence

Project Overview

What if your 500-page technical manual could answer questions? At American Axle Manufacturing, I built exactly that — a conversational AI that transforms static PDFs, CSVs, and DOCX files into interactive knowledge assets.

Built on LangChain + FAISS + Cohere in just 8 weeks (research → production!), this RAG system reduced information retrieval time from 45 minutes to under 2 minutes.

The Problem

Enterprise documentation is a graveyard of underutilized knowledge:

• Engineers spend 20-45 minutes searching through manuals
• Technical specs exist but aren't accessible to non-technical users
• Knowledge walks out the door when SMEs leave
• Compliance audits require digging through mountains of PDFs

The solution? Make documents conversational.

Key Features

• Multi-Format Ingestion: PDF, CSV, DOCX support with intelligent text extraction and structure preservation

• Semantic Chunking: 500-1000 token segments with overlap, balancing context richness and retrieval granularity

• FAISS Vector Search: Sub-second similarity matching across document collections using Cohere embeddings

• Conversational Memory: Multi-turn interactions with LangChain's ConversationalRetrievalChain maintaining context across questions

• Source Attribution: Every answer links back to specific document passages — perfect for compliance and verification

Technical Architecture

User Upload → Document Parsing → Semantic Chunking
                                       ↓
                              Cohere Embeddings
                                       ↓
                              FAISS Vector Index
                                       ↓
User Query → Query Embedding → Similarity Search → Top-K Retrieval
                                                          ↓
                                              LLM Answer Generation (Cohere)
                                                          ↓
                                              Response + Source Attribution

RAG Pipeline Details

Embedding Generation

• Cohere embed-english-v3.0 (768-dimensional vectors)
• Batched API calls with retry logic
• Metadata association for traceability

Vector Store

• FAISS Flat index for accurate retrieval (perfect for mid-sized corpora)
• Persistence to disk — no re-embedding needed
• Index lifecycle management for document updates

Conversational Chain

• LangChain ConversationalRetrievalChain with memory buffer
• Carefully crafted prompts emphasizing document grounding
• Temperature tuning for consistent, factual responses

Technology Selection Rationale

Why Cohere?

• Enterprise-grade reliability and stable API contracts
• Strong instruction-following for document Q&A
• Cost-effective for mid-to-large document collections
• Official LangChain integration with comprehensive docs

Why FAISS?

• Production-proven at Facebook scale
• CPU-friendly with fast similarity search
• No external dependencies (unlike Qdrant, Milvus)
• Seamless LangChain abstraction for future migration

Why LangChain?

• Pre-built RAG chains reduce implementation time
• Modular design enables easy component swapping
• Prompt engineering best practices baked in
• Strong community and ecosystem

8-Week Sprint Timeline

Week	Phase	Deliverable
1-2	Research	Tech selection matrix, architecture design
3-4	Core Pipeline	PDF parsing, embeddings, FAISS indexing
5-6	RAG Integration	Full chain wired with Streamlit UI
7-8	Validation	Testing with real docs, latency optimization

Results

• Indexing: 50+ technical documents in <5 minutes
• Query Latency: <3 seconds for typical questions
• User Feedback: Positive ratings from engineers, managers, support staff
• Information Retrieval: 45 minutes → 2 minutes (95%+ reduction)

Business Impact

For Engineers: Query specs, test reports, work instructions in seconds For Managers: Ask targeted questions about KPIs, metrics, program status For New Hires: Onboard faster by exploring documentation interactively For Compliance: Audit trails via source attribution

Tech Stack

AI/LLM: Cohere Chat/Command, Cohere Embeddings Framework: LangChain, FAISS Backend: Python 3.10+ Frontend: Streamlit Parsing: PyPDF2, python-docx

Key Learnings

This project taught me the RAG pattern from scratch — not just implementation, but the critical decisions:

• Chunk size matters enormously (too small = lost context, too large = retrieval noise)
• Top-K selection is an art (3-5 works well for generation)
• Prompt engineering is underrated (the difference between hallucination and grounded answers)

Most importantly: 8 weeks is enough to go from zero to production if you make decisive technology choices early.

---

My first production AI system — and the one that convinced me that LLMs aren't just chatbots, they're interfaces to human knowledge.