🏭 Industrial RAG Assistant: Siemens S7-1200 Expert

An enterprise-grade technical assistant utilizing Retrieval-Augmented Generation (RAG) to query 1,500+ pages of the Siemens S7-1200 System Manual.

💎 Core Capabilities

• Semantic Retrieval: Leverages vector embeddings to interpret technical intent beyond simple keyword matching.
• Audit-Ready Citations: Every response is mapped to the exact page number of the source PDF for technical verification.
• Industrial Safety Focus: Designed to minimize "search latency" for field engineers while providing direct links to official documentation.

💻 Technical Infrastructure

• LLM: Google Gemini 2.5 Flash
• Vector DB: ChromaDB
• Framework: LangChain Classic
• UI: Streamlit

🛠️ Deployment Guide

1. Clone Repository: git clone https://github.com/alirezasoroushe/industrial-rag-assistant.git
2. Environment Configuration: pip install -r requirements.txt
3. Authentication: .env with your GOOGLE_API_KEY.
4. Initialize Application: streamlit run app.py

🛠️ Project Troubleshooting & Industrial Insights

Building this system in the fast-evolving AI landscape presented several "real-world" challenges. Below are the technical hurdles overcome and the "Pro-Tips" for deploying RAG in industrial environments.

1. Handling API Deprecations (The "404 NOT_FOUND" Challenge)

Problem: Initially, the project used gemini-1.5-flash. However, as of early 2026, Google retired these legacy identifiers, leading to 404 errors in the v1beta API.
Solution: Migrated the logic to the Gemini 2.5 Flash stable release.
Tip: Always hard-code specific model versions (e.g., gemini-2.5-flash) rather than using -latest tags in production to prevent breaking changes when models are updated.

2. LangChain Modularization (ModuleNotFoundError)

Problem: LangChain split its core logic from community integrations, causing imports like RetrievalQA to fail.
Solution: Installed langchain-classic and updated import paths to langchain_classic.chains to bridge the legacy logic with the new 2026 framework.
Tip: Use a virtual environment (Conda or Venv) to isolate dependencies, ensuring that "Community" packages don't conflict with "Core" AI logic.

3. Precision Over Performance

Problem: Standard semantic search sometimes missed specific Siemens part numbers or error codes.
Solution: Implemented a 1000-character chunk size with a 200-character overlap. This ensures that technical specs aren't "cut in half," maintaining the context of safety warnings.
Tip: In RAG systems for manuals, Metadata is King. By preserving the page number in the metadata, we provide a "human-in-the-loop" verification step, essential for industrial safety standards.

4. Embedding Model Synchronization

Problem: Using a 2024 embedding model with a 2026 LLM caused "dimensionality mismatch" errors.
Solution: Standardized the entire pipeline on the text-embedding-004 engine to ensure the "Brain" and the "Memory" speak the same mathematical language.

🏗️ System Architecture

graph TD
    subgraph "Knowledge Ingestion"
        A["Siemens PDF Manual"] --> B["PyPDFLoader"]
        B --> C["RecursiveTextSplitter"]
        C --> D["Chroma Vector DB"]
    end

    subgraph "Retrieval-Augmented Generation"
        direction TB
        %% Invisible spacer to move title up
        spacer1[ ] --- User
        style spacer1 fill:none,stroke:none,color:#00000000
        
        User["User Query"] --> Embed["Google text-embedding-004"]
        Embed --> Search["Vector Search"]
        D -.-> Search
        Search --> Context["Relevant Context + Citations"]
    end

    subgraph "Final Response"
        Context --> Gemini["Gemini 2.5 Flash"]
        User --> Gemini
        Gemini --> Response["Verified Technical Answer"]
    end

    style D fill:#f9f,stroke:#333,stroke-width:2px
    style Gemini fill:#00f,color:#fff

Loading