Link a Langfuse generation to the exact prompt version that produced it, enabling cost and quality comparison across prompt versions

Documented steps

1. Fetch the prompt object first rather than hardcoding text, so you have a reference to attach: Python `prompt = langfuse.get_prompt("movie-critic")`; JS/TS `const prompt = await langfuse.prompt.get("movie-critic")`.
2. In Python with the decorator or context-manager pattern, pass `prompt=prompt` when starting or updating the generation: `langfuse.update_current_generation(prompt=prompt)`, or `with langfuse.start_as_current_observation(as_type="generation", prompt=prompt) as generation: ...`.
3. For manual (non-context-manager) observations, pass it at creation: `generation = langfuse.start_observation(as_type="generation", prompt=prompt)`.
4. In JS/TS, attach it via `updateActiveObservation({ prompt }, { asType: "generation" })`, `generation.update({ prompt })`, or at creation with `startObservation("llm", { prompt }, { asType: "generation" })`.
5. If one prompt drives several generations in a single request (multi-step chains), use Python's `propagate_attributes(prompt=prompt)` context manager so every nested generation inherits the link without repeating the parameter.
6. For the OpenAI integration, pass the prompt object as `langfuse_prompt=prompt` in `openai.chat.completions.create(...)`. For LangChain, set `metadata={"langfuse_prompt": langfuse_text_prompt}` on the PromptTemplate. For the Vercel AI SDK, pass `metadata: { langfusePrompt: fetchedPrompt }`.
7. Open a trace in the UI and confirm the linked prompt name and version badge appear, then use the Prompts view metrics panel (or filter traces by prompt name/version) to compare cost, latency, and scores across versions.
8. Verified against current official docs on 2026-07-29: https://langfuse.com/docs/prompt-management/features/link-to-traces , https://langfuse.com/docs/prompt-management/get-started