ferro-labs/ai-gateway

Summary The gateway fires EventHookFunc hooks on every completed or failed request (gateway.go), but there is no built-in way to forward those events to an external system (e.g. a logging aggregator, analytics pipeline, or audit store) without writing custom Go code. A webhook output plugin would let operators configure one or more HTTP endpoints to receive gateway events as JSON payloads — no code required. Proposed design Plugin identifier: "webhook" Config example plugins: - name: webhook stage: after_request config: url: https://ingest.example.com/events headers: Authorization: Bearer ${WEBHOOK_SECRET} timeout_seconds: 5 retry_attempts: 2 on_error: log # or "drop" — never block the hot path Payload (JSON) { "trace_id": "abc123", "model": "gpt-4o", "provider": "openai", "prompt_tokens": 120, "completion_tokens": 340, "latency_ms": 812, "error": null, "timestamp": "2026-03-12T10:00:00Z" } Behaviour The webhook cal

Summary The existing response-cache plugin uses exact-match hashing. This is useful for identical repeated requests but misses the common case where two prompts are semantically equivalent (e.g. "What is the capital of France?" vs "Tell me the capital city of France"). A semantic cache plugin would embed incoming prompts, compare against cached embeddings by cosine similarity, and serve cached responses when the similarity exceeds a configurable threshold — drastically reducing provider calls and cost for FAQ-style workloads. Proposed design Plugin identifier: "semantic-cache" Config options plugins: - name: semantic-cache stage: before_request config: similarity_threshold: 0.92 # cosine similarity (0.0–1.0) max_entries: 500 max_age: 600 # seconds embedding_provider: openai # which gateway provider to use for embeddings embedding_model: text-embedding-3-small Behaviour Before request: embed the user turn, scan the