LLM Restaurant Finder
LLM Restaurant Finder turns conversational prompts into structured Foursquare searches via a TypeScript Hono API — Gemini drafts the JSON intent, Zod validates it, and the mapper hits Places.
A lightweight Hono route receives user prompts and metadata via a JSON payload. The handler trims whitespace, assigns default geolocation hints when a user omits them, and tags each request with a correlation ID that threads through the rest of the pipeline for observability.
The sanitized prompt is embedded into a Gemini-provided system prompt that instructs the model to answer strictly with JSON matching a Zod schema. Gemini responses are parsed back into TypeScript, validated again on the server, and any deviations trigger a user-facing remediation message instead of continuing with unreliable data.
Valid JSON commands hydrate a mapper that translates abstract cuisine, mood, and budget directives into concrete Foursquare filter sets. Localized price tiers, radius units, and category IDs are all resolved here so the Places API receives deterministic parameters.
The primary request hits the Foursquare Places search endpoint to gather a candidate list. When Gemini flags the need for ambience or operating-hours insights, the service fans out detail calls per place, caching responses to stay within rate limits.
Aggregated data is flattened into a uniform card model with consistent distance, rating, and address fields. Zod schemas enforce that partial responses still surface predictable keys, enabling a resilient React presentation layer.
Before any Gemini tokens are spent, middleware checks for a rotating access code stored in environment variables. Requests missing or using outdated codes receive a 401 response, shielding both the LLM and Foursquare quotas from untrusted traffic.
The API is deployed on Render using the Node.js 20 runtime. Build hooks install dependencies, run linting, and then start the Hono server as a long-lived process with environment variables providing Gemini and Foursquare credentials.
- Gemini occasionally over-specifies filters, leading to sparse results in smaller cities; fallback prompts ask users to broaden the query.
- Foursquare Places API quotas cap burst throughput, so enrichment is throttled when rate limits approach their ceiling.
- Access-code distribution remains manual, which can slow onboarding for collaborators.