Capture an LLM generation call (prompt, completion, tokens, latency) as a PostHog LLM analytics event

Documented steps

1. Use event name $ai_generation, sent like any other event to the public capture endpoint: POST https://us.i.posthog.com/i/v0/e/ (or eu.i.posthog.com) with Content-Type: application/json.
2. Body: api_key (project token), event: "$ai_generation", distinct_id, properties: { $ai_trace_id, $ai_model, $ai_provider, $ai_input, $ai_output_choices, $ai_input_tokens, $ai_output_tokens, $ai_latency }.
3. $ai_trace_id groups related AI events (all steps of one LLM session) and must contain only letters, numbers, and the characters - _ ~ . @ ( ) ! ' : | — a UUID works.
4. $ai_model is the model used (e.g. "gpt-4o"), $ai_provider the vendor (e.g. "openai", "anthropic"); $ai_input is the message list sent (with role fields) and $ai_output_choices the response choices.
5. $ai_input_tokens / $ai_output_tokens capture token counts; $ai_latency (seconds) is optional but recommended.
6. Docs: https://posthog.com/docs/llm-analytics/installation/manual-capture