An intelligent document search system powered by agentic AI search. Uses autonomous multi-step reasoning to search, analyze, and synthesize information from document collections using Cloudflare Workers, AI Search, and Google Gemini.
- 🤖 Agentic AI Search - Autonomous multi-step reasoning with Google Gemini
- 🔍 Semantic Search - Powered by Cloudflare AI Search
- ✏️ Contextual Query Rewriting - Automatically consolidates follow-up questions into context-aware queries
- 🧠 Knowledge Synthesis - Iteratively gathers and evaluates information
- 💬 Modern Chat Interface - Built with vanilla HTML, CSS (Tailwind), and jQuery
- ⚡ Edge Deployment - Runs on Cloudflare Workers
- 💾 Persistent Sessions - Room-based conversation history with automatic loading on reconnect
- 🔄 Multi-RAG Support - Switch between multiple RAG instances via dropdown (creates new room automatically)
npm installGet your Google API key from Google AI Studio
For local development:
Create a .dev.vars file in the root directory:
cp .dev.vars.example .dev.varsThen edit .dev.vars and add your API key:
GOOGLE_GENERATIVE_AI_API_KEY=your_actual_api_key_here
For production deployment: Set the secret using Wrangler:
npx wrangler secret put GOOGLE_GENERATIVE_AI_API_KEYOr set it in the Cloudflare dashboard under Workers > Your Worker > Settings > Variables
The application uses AI Search instances for document retrieval. Make sure:
- You have one or more AI Search instances in your Cloudflare account
- Your documents are indexed in the AI Search instances
- The AI binding is configured in
wrangler.jsonc(already done) - The default RAG instance in
src/index.tscan be changed if needed (currentlypublic-reports)
The UI automatically detects and lists all available RAG instances in a dropdown menu.
The application uses persistent room-based sessions with the following behavior:
- Room Persistence: Each chat session has a unique room ID stored in localStorage
- Message History: When reconnecting to an existing room, all previous messages are automatically loaded and displayed
- RAG Switching: Changing the RAG instance automatically creates a new room since different RAGs contain different document collections
- New Chat: Click "New Chat" button to start a fresh conversation in a new room while preserving the previous room's history
Run locally:
npm run devThe chat interface will be available at http://localhost:8787
Deploy to Cloudflare Workers:
npm run deploy- User Query - User asks a question through the chat interface
- Query Rewriting - If there are previous user queries in the session, Gemini 2.5 Flash consolidates them into a single, contextually-aware search query
- Agentic Loop - Autonomous reasoning loop with up to 5 iterations
- Search & Extract - Agent searches documents and extracts relevant knowledge
- Evaluate Sufficiency - Structured decision on whether more information is needed
- Iterate or Complete - Continue with new search queries or finalize answer
- Synthesize Response - Agent generates comprehensive answer from accumulated knowledge
- Persist Session - Conversation saved to room-specific Durable Object storage
- Frontend: Vanilla HTML, Tailwind CSS, jQuery
- Backend: Cloudflare Workers (TypeScript)
- AI Model: Google Gemini 2.5 Pro & Flash
- AI SDK: Vercel AI SDK with Agents SDK
- Vector Search: Cloudflare AI Search
- Storage: Cloudflare Durable Objects
- Deployment: Cloudflare Workers
Lists all available RAG instances.
Response:
{
"success": true,
"rags": [
{
"id": "public-reports",
"enable": true,
"type": "r2",
"source": "gov-public-dataset",
"paused": false,
"status": "waiting",
...
}
]
}Main chat endpoint for AI agent interactions.
Request:
{
"message": "What information is available about...?"
}Response:
{
"message": "Based on the search results...",
"sources": [
{
"file_id": "doc123",
"filename": "healthcare-policy-2020.pdf",
"score": 0.85,
"content": [...]
}
],
"steps": 2
}flowchart TD
Start([User Query via WebSocket]) --> LoadMessages[Load Persisted Messages<br/>from Durable Object]
LoadMessages --> CheckHistory{Has Previous<br/>User Messages?}
CheckHistory -->|Yes| QueryRewrite[Query Rewriting<br/>Gemini 2.5 Flash<br/>Consolidate context]
CheckHistory -->|No| InitLoop[Initialize Agentic Loop<br/>maxIterations = 5]
QueryRewrite --> InitLoop
InitLoop --> LoopStart{Start Iteration<br/>iteration++}
LoopStart --> KnowledgeGen[Knowledge Generation<br/>Gemini 2.5 Flash]
KnowledgeGen --> SearchTool[searchDocuments Tool<br/>max 3 tool calls per iteration]
SearchTool --> AISearch[Cloudflare AI Search<br/>Search RAG Instance]
AISearch --> SearchResults[Search Results<br/>max 10 documents]
SearchResults --> ExtractKnowledge[Extract 3-5 Key Facts<br/>with Document References]
ExtractKnowledge --> Accumulate[Accumulate Knowledge<br/>Add to Knowledge List]
Accumulate --> Evaluate[Evaluate Sufficiency<br/>Gemini 2.5 Flash]
Evaluate --> Decision{Structured Output:<br/>isKnowledgeEnough?}
Decision -->|No Knowledge| ExitLoop[Exit Loop]
Decision -->|Yes & < 5 iterations| GenerateQuery[Generate Next<br/>Search Query]
Decision -->|Sufficient Knowledge| ExitLoop
Decision -->|Max Iterations Reached| ExitLoop
GenerateQuery --> LoopStart
ExitLoop --> Synthesize[Synthesize Final Answer<br/>Gemini 2.5 Pro]
Synthesize --> StreamResponse[Stream Response<br/>Text Deltas via WebSocket]
StreamResponse --> AddToHistory[Add to Message History]
AddToHistory --> PersistMessages[Persist Messages to<br/>Durable Object Storage]
PersistMessages --> SendFiles[Send File Metadata<br/>Download Links]
SendFiles --> End([Complete])
style Start fill:#e1f5ff
style End fill:#e1f5ff
style QueryRewrite fill:#e1e8ff
style KnowledgeGen fill:#fff4e1
style Evaluate fill:#fff4e1
style Synthesize fill:#ffe1f5
style AISearch fill:#e1ffe1
style Decision fill:#ffe1e1
style PersistMessages fill:#f0e1ff
Frontend (WebSocket Client)
- Vanilla HTML, Tailwind CSS, jQuery
- Real-time bi-directional communication
- Multi-RAG instance selector
- Session-based chat rooms
Backend (Cloudflare Workers)
SearchAgent- Durable Object class extendingAIChatAgent- Persistent message storage per room
- Multi-step agentic reasoning loop
- Document retrieval and streaming
Query Rewriting Phase
- Triggered when user sends follow-up queries in the same session
- Consolidates all previous user queries into context-aware search query
- Handles conversational refinements (e.g., "and without a tire" → "can a car run without a tire")
- Maintains conversation continuity while keeping searches focused
Agentic Loop Process
-
Knowledge Generation Phase (Gemini Flash)
- Uses
searchDocumentstool to query Cloudflare AI Search - Can perform up to 3 searches per iteration
- Extracts 3-5 key knowledge entries from results
- Accumulates knowledge across iterations
- Uses
-
Evaluation Phase (Gemini Flash)
- Structured output with Zod schema
- Determines if accumulated knowledge is sufficient
- Generates next search query if more info needed
- Considers query coverage, gaps, and context
-
Stopping Conditions
- Knowledge deemed sufficient
- Maximum 5 iterations reached
- No new knowledge extracted
- No next search query generated
-
Synthesis Phase (Gemini Pro)
- Combines all accumulated knowledge
- Generates comprehensive, structured answer
- Streams response in real-time
- Excludes document citations (links sent separately)
Storage & Persistence
- Durable Objects for stateful WebSocket connections
- Room-based message history
- Automatic state synchronization
- File metadata tracking
GOOGLE_GENERATIVE_AI_API_KEY- Your Google AI API key (required)
MIT