[FEAT]: Automated External Dataset (RAG) connection for instant lookup of protocols and rules

[Cubix33]

📝 Description

Currently, the AI only knows what is spoken in the audio transcript. This feature gives the AI an "open-book test" capability using Retrieval-Augmented Generation (RAG).

If a transcript mentions a specific situation (like a "Class 3 Chemical Leak"), the AI will:

1. Automatically search local fire department manuals or HAZMAT guidelines.
2. Retrieve the exact safety protocol.
3. Attach that official protocol directly to the final report's notes.

💡 Rationale

• Instant Expert Knowledge: First responders cannot memorize every single hazardous material code or standard operating procedure (SOP). FireForm will act as an intelligent incident partner, not just a form filler.
• Zero Hallucinations: Grounding the AI in factual, local documents prevents it from making up fake safety rules.
• Infinite Scalability: Any fire department can customize FireForm just by dropping their own PDF manuals into the database.

🛠️ Proposed solution

I will implement a lightweight RAG pipeline to give the LLM access to external knowledge.

• Database Setup: Integrate a local vector database (like ChromaDB or FAISS) to store embeddings of safety manuals.
• Context Retrieval (src/llm.py): Add a pre-processing step to scan the transcript, search the vector database, and grab the most relevant chunks of text.
• Prompt Injection (src/llm.py): Update the build_prompt function to feed this new "Official Context" directly to the Mistral model so it can supplement its answers.

✅ Acceptance Criteria

• Ingestion: The system successfully converts a sample PDF manual into searchable vector embeddings.
• Retrieval: If the audio transcript mentions "Ammonia Spill," the final JSON output successfully includes the Ammonia handling protocol from the manual.
• Stability: The new vector database runs perfectly inside the existing Docker container without breaking the build.