confidential-websearch

README

Confidential Web Search Proxy

A proxy that augments LLM responses with real-time web search results. It uses a two-model architecture:

1. Agent Model (small, fast, configured): Decides whether to search and what to search for
2. Responder Model (from request): Generates the final response using search results

Uses the Tinfoil Go SDK for secure, attested communication with Tinfoil enclaves.

Architecture

User Request
  │ model: "kimi-k2-thinking"
  │ Authorization: Bearer <api-key>
  ▼
┌─────────────────┐
│   Agent Model   │ ──► Decides: search needed?
│  (gpt-oss-120b) │     (configured, small/fast)
└─────────────────┘
         │
         ▼ web_search tool call
┌────────────────┐
│  Exa / Bing    │ ──► Returns search results
└────────────────┘
         │
         ▼
┌─────────────────────────────────────┐
│         Responder Model             │
│     (from request: kimi-k2-thinking)│ ──► Generates response
└─────────────────────────────────────┘
         │
         ▼
   Final Response (streamed)

Quick Start

# Set a search API key (Exa preferred, Bing as fallback)
export EXA_API_KEY="your-exa-api-key"
# OR
export BING_API_KEY="your-bing-api-key"

# Run the proxy
go run .

Environment Variables

Variable	Default	Description
EXA_API_KEY	-	Exa AI Search API key (preferred)
BING_API_KEY	-	Bing Search API key (fallback)
AGENT_MODEL	gpt-oss-120b	Agent model for tool use decisions
LISTEN_ADDR	:8089	Address to listen on

The proxy will use Exa if EXA_API_KEY is set, otherwise falls back to Bing.

API

The proxy exposes an OpenAI-compatible /v1/chat/completions endpoint:

• The model field specifies which model generates the final response
• The Authorization header is forwarded to backend models

curl http://localhost:8089/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $TINFOIL_API_KEY" \
  -d '{
    "model": "kimi-k2-thinking",
    "messages": [{"role": "user", "content": "What is the latest news about SpaceX?"}],
    "stream": true
  }'

How It Works

1. Request arrives with model and API key
2. Agent phase: Small agent model (gpt-oss-120b) decides if search is needed
   • If yes, generates search queries using the web_search tool
   • Searches are executed in parallel via Exa or Bing
3. Response phase: Search results are injected into the context as tool results
4. Final response: The model from the request generates the answer
5. Response is streamed back to the client

Docker

docker build -t websearch-proxy .
docker run -p 8089:8089 \
  -e EXA_API_KEY=$EXA_API_KEY \
  websearch-proxy

Security

This proxy uses the Tinfoil Go SDK which provides:

• Automatic attestation validation to ensure enclave integrity
• TLS certificate pinning with attested certificates
• Direct-to-enclave encrypted communication
• API key forwarding from client requests (no stored credentials)