LLMonFHIR Firebase Backend

Firebase Functions backend for LLMonFHIR with a RAG-enabled OpenAI-compatible /chat endpoint and a lightweight web client for comparison testing.

Quickstart

Backend (emulators)

Note
functions/.secret.local must contain a valid OPENAI_API_KEY

cd functions
npm install
cd ..
sh run-emulator.sh

This installs backend dependencies, builds functions, and starts Firebase emulators. Ensure the OPENAI_API_KEY secret is configured in your Firebase project before deploying.

Web client (optional)

cd web
npm install
npm run dev

The web UI compares responses with RAG enabled and disabled using mock FHIR tool results.

Architecture

Realtime chat path for LLMonFHIR / web client
  → OpenAI-compatible Firebase Function `/chat`
      (drop-in replacement for OpenAI `/v1/chat/completions`)
      (client keeps the same request body; only the URL changes)
      → Retrieve top-k chunks (Genkit retriever)
          → Vector store (dev-local)
      → Augment system prompt
  ← Streamed response (+ optional RAG metadata)

Document ingestion path for Firebase Storage `rag_files/*.pdf`
  → `onPDFUploaded` trigger
      → PDF text extraction
      → Chunk + embed
      → Index into vector store

Core Components

OpenAI-compatible chat proxy

• File: functions/src/functionImplementations/openai-proxy.ts
• Firebase Function name: chat
• Endpoint: https://<region>-<project>.cloudfunctions.net/chat
• Drop-in OpenAI /v1/chat/completions replacement: keep the same request body and only change the URL
• Supports streaming and non-streaming OpenAI chat requests
• Injects retrieved RAG context into the system prompt
• Toggle RAG off for debugging: ?ragEnabled=false

PDF indexing via storage trigger

• File: functions/src/functionImplementations/storage-trigger.ts
• Trigger: new PDF uploaded under rag_files/ in Storage
• Extracts text, cleans and chunks it, and embeds content into the vector store

RAG pipeline internals

• Chunking: functions/src/rag/chunker.ts
• Indexing: functions/src/rag/indexer.ts
• Retrieval: functions/src/rag/retriever.ts (top 5 chunks)
• Genkit config: functions/src/utils/genkit.ts

Backend Setup

Install and build

cd functions
npm install
npm run build

Emulators

sh run-emulator.sh

Deploy

firebase deploy --only functions

Configure secrets

firebase functions:secrets:set OPENAI_API_KEY

API Usage

Request (OpenAI-compatible)

curl -X POST "https://<firebase-functions-url>/chat" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [
      { "role": "user", "content": "Summarize the latest lab results." }
    ],
    "stream": true
  }'

Streaming response (RAG metadata)

When RAG is enabled and context is found, the stream includes a metadata event before the usual OpenAI deltas:

{
  "type": "rag_context",
  "context": "[Document: ...]",
  "contextLength": 1234,
  "enabled": true
}

Non-streaming response (RAG metadata)

The non-streaming response includes _ragContext:

{
  "id": "...",
  "choices": [ ... ],
  "_ragContext": {
    "context": "[Document: ...]",
    "contextLength": 1234,
    "enabled": true
  }
}

Web Client

The /web directory contains a minimal React app that compares responses with RAG enabled and disabled. It routes OpenAI SDK calls to the Firebase Functions /chat endpoint and uses mock FHIR tool outputs to simulate data retrieval.

Limitations and Next Steps

• The RAG vector store uses @genkit-ai/dev-local-vectorstore for development only.
• For production, replace it with a persistent vector store such as Firestore Vector Search: https://genkit.dev/docs/integrations/cloud-firestore/