langfuse

func AsynqMiddleware ¶

func AsynqMiddleware() asynq.MiddlewareFunc

AsynqMiddleware is the worker-side counterpart of GinMiddleware. It:

1. Extracts the W3C traceparent stamped onto the task payload by InjectTracing and resumes the originating trace (so the Langfuse UI stitches the HTTP request and the async processing into one tree). For scheduled jobs with no upstream trace it opens a standalone trace named after the task type.

2. Opens a SPAN around the handler execution so every child generation (embedding / VLM / chat / rerank / ASR) auto-attaches to it.

3. Enriches the span with asynq's own metadata: task id, queue, retry count, payload size.

When the manager is disabled it degrades to a pass-through; failure of the Langfuse path never blocks task execution. Register once via mux.Use.

func AttachTraceparent ¶

func AttachTraceparent(ctx context.Context, traceparent string) context.Context

AttachTraceparent resumes the originating trace from a W3C traceparent previously captured by TraceparentFromContext / InjectTracing. When ctx already has a *Trace (same-request background work), it is left unchanged so we do not replace a live local parent with a remote one.

This is the in-process counterpart of AsynqMiddleware's extract path: a later HTTP request (e.g. POST .../suggestions) has no GinMiddleware trace, and without this StartGeneration would auto-create an orphan root named after the LLM call.

func GinMiddleware ¶

func GinMiddleware() gin.HandlerFunc

GinMiddleware returns a Gin handler that opens a Langfuse trace for each incoming request that hits a traced path. The trace is auto-finished when the handler chain returns; individual LLM calls inside the handler attach their generations to this trace via the request context.

Only paths matching shouldTrace are traced — static assets, health checks and polling endpoints are noisy and uninteresting.

func InjectTracing ¶

func InjectTracing(ctx context.Context, carrier types.LangfuseTracingCarrier)

InjectTracing stamps the current W3C traceparent (plus a best-effort user/session label) from ctx onto the given payload, provided the payload embeds types.TracingContext (and thus implements LangfuseTracingCarrier).

The traceparent is produced by propagator.Inject from the active OTel span context (the HTTP root span opened by GinMiddleware). The asynq worker re-extracts it so worker-side spans are children of the same trace — giving LiteFuse one stitched tree across the HTTP request and the async processing. This also makes a sop3 run's traceparent propagate through to any asynq jobs WeKnora enqueues while serving sop3's agent-chat call.

Safe to call unconditionally: when Langfuse is disabled or no span is present on ctx, it writes a zero-valued TracingContext — which round-trips through JSON as absent fields and costs nothing.

func SummarizeMemoryRecallOutput ¶

func SummarizeMemoryRecallOutput(meta map[string]interface{}, items []*types.MemoryItem) map[string]interface{}

SummarizeMemoryRecallOutput builds Langfuse output for a memory.recall span.

func SummarizePassagePreviews ¶

func SummarizePassagePreviews(
	candidates []*types.SearchResult,
	passages []string,
	limit int,
) []map[string]interface{}

SummarizePassagePreviews builds rerank passage previews aligned with candidates.

func SummarizeRankScores ¶

func SummarizeRankScores(
	results []map[string]interface{},
	limit int,
) []map[string]interface{}

SummarizeRankScores builds rerank model score rows for Langfuse output.

func SummarizeRetrievalContextOutput ¶

func SummarizeRetrievalContextOutput(
	background string,
	interests, documents []string,
	items []*types.MemoryItem,
) map[string]interface{}

SummarizeRetrievalContextOutput builds Langfuse output for retrieval conditioning.

func SummarizeRetrieveOutput ¶

func SummarizeRetrieveOutput(results []*types.RetrieveResult) map[string]interface{}

SummarizeRetrieveOutput builds Langfuse output for the retrieve span.

func SummarizeSearchResults ¶

func SummarizeSearchResults(results []*types.SearchResult, limit int) map[string]interface{}

SummarizeSearchResults builds a compact ranked preview for rerank spans.

func TraceparentFromContext ¶

func TraceparentFromContext(ctx context.Context) string

TraceparentFromContext returns the W3C traceparent for the active span, or empty when Langfuse is disabled or ctx carries no span. Used to stamp the originating chat trace onto derived work (follow-up suggestions) that may run after the HTTP handler returns, or on a later request.

func TruncateRunes ¶

func TruncateRunes(s string, maxRunes int) string

TruncateRunes shortens s to at most maxRunes runes for trace payloads.