Route LLM requests with Anthropic, OpenAI, Gemini, Postgres, and Slack

Quick overview This workflow exposes a POST webhook that sends a prompt through a resilient LLM fallback chain (Anthropic, OpenAI, then Google Gemini), validates outputs against deterministic rules/JSON schema, logs all attempts to Postgres with a circuit breaker, and posts a Slack alert when failover or degraded responses occur.

How it works Receives a POST request on the n8n webhook (or runs manually for testing) with an LLM prompt and optional validation settings. Normalizes the request by applying defaults, coercing types, and generating a prompt hash used for stale-cache fallback. Reads provider health from Postgres and builds an ordered provider chain (optionally honoring a preferred provider and skipping providers with an open circuit). Calls Anthropic, OpenAI, or Google Gemini via HTTP, then normalizes the response and validates it (status codes, empty/truncated/blocked/refusal output, length/keyword rules, and optional JSON parsing plus schema checks). Logs each attempt to Postgres and updates circuit-breaker state, then either retries with backoff/repair hints, switches providers, or exhausts the chain. If all providers fail, looks up a recent successful response for the same prompt hash in Postgres and returns it, otherwise returns a static degraded message with HTTP 503. Saves the final result to Postgres, posts a Slack alert only when fallback/degraded mode is used, and responds to the webhook with the final JSON payload.

Setup Create HTTP Header Auth credentials for Anthropic (x-api-key), OpenAI (Authorization: Bearer …), and Google Gemini (x-goog-api-key) and attach them to the three HTTP Request nodes. Add a Postgres credential, run the provided database schema (tables for llm_requests, llm_attempts, and llm_provider_health), and confirm the workflow’s queries match your schema name. Add a Slack credential and set the target channel name in the configuration (for example, #llm-alerts). Review and update the fallback policy in the configuration values (provider order, models, timeouts, retry/backoff, circuit breaker thresholds, cache window, and degraded message). Copy the webhook URL from the Webhook trigger and configure your calling app to send POST requests to /llm-fallback-chain with the expected JSON payload.