Langfuse Extract

Langfuse Extract

Read-side tool: pulls data OUT of Langfuse (traces, sessions, observations/generations, prompts, scores, daily metrics) via the public REST API. For instrumenting an app to SEND traces to Langfuse, use the langfuse-observability skill instead.

When to Activate

• "dame las traces de ayer con error"
• "extrae las sessions del usuario X"
• "necesito el historial de prompts de support_agent_v2"
• "cuánto costó el proyecto Y la semana pasada" (daily metrics)
• Building an eval/fine-tuning dataset from production traces
• Auditing what a specific session/trace actually sent to the model

Credentials — .env convention

Same pattern as ~/.intercom_*.env / ~/.twilio_*.env: one .env file per workspace/project at ~/.langfuse_<workspace>.env, never committed to git.

Template: .env.example in this skill's directory. Required vars:

WORKSPACE_NAME=       # label for logs
LANGFUSE_HOST=        # https://cloud.langfuse.com (EU) / https://us.cloud.langfuse.com (US) / self-hosted
LANGFUSE_PUBLIC_KEY=  # pk-lf-...
LANGFUSE_SECRET_KEY=  # sk-lf-...

Get keys from the Langfuse project: Settings > API Keys. Auth is HTTP Basic — username=public key, password=secret key. Region matters: a US cloud project's keys will 401 against the EU host and vice versa.

Set chmod 600 on the env file.

Tools in this skill

• scripts/langfuse_client.py — reusable LangfuseClient class (auth, pagination, retry on 429/5xx). Import this if writing a custom extraction script.
• scripts/extract.py — CLI wrapper. Covers the common cases without writing code.

CLI usage

cd ~/.claude/skills/langfuse-extract/scripts
python extract.py <kind> --env-file ~/.langfuse_prod.env [options]

<kind>: traces | trace | sessions | session | observations | prompts | prompt | scores | metrics | health

# Traces in a date range, filtered by session
python extract.py traces --env-file ~/.langfuse_prod.env \
    --from 2026-07-01 --to 2026-07-16 --session-id abc123 -o traces.jsonl -f jsonl

# Full session detail (includes its traces)
python extract.py session --env-file ~/.langfuse_prod.env --session-id abc123

# Debugging shortcut: "what changed in this agent's prompt" for a whole session —
# every GENERATION across all its traces, with model + linked prompt name/version + input/output
python extract.py debug-session --env-file ~/.langfuse_prod.env --session-id abc123

# Generations only (skip spans/events) for a trace — useful to see exact model I/O
python extract.py observations --env-file ~/.langfuse_prod.env \
    --trace-id abc123 --type GENERATION

# Specific prompt version/label
python extract.py prompt --env-file ~/.langfuse_prod.env \
    --name support_agent_v2 --label production

# Scores (evals / human feedback) in a window
python extract.py scores --env-file ~/.langfuse_prod.env --from 2026-07-01

# Daily cost/usage metrics for one traced function
python extract.py metrics --env-file ~/.langfuse_prod.env \
    --trace-name generate_reply --from 2026-07-01

# Sanity check credentials/host before a big pull
python extract.py health --env-file ~/.langfuse_prod.env

Output: --output/-o file --format/-f json|jsonl|csv (default: prints JSON to stdout). csv flattens nested fields (input/output/metadata) to JSON strings per cell.

--max-items N (default 500) caps TOTAL items across all pages for traces, observations, scores — not just page size (--limit is page size only). Without this cap an unscoped list call on an active project pages forever. session/trace/prompt (single-resource get) and debug-session (bounded by one session's traces) ignore it — there's nothing to cap.

API reference (what the client wraps)

Base path: {LANGFUSE_HOST}/api/public/.... All list endpoints are paginated (page, limit params; response { data: [...], meta: { page, totalPages, totalItems } }) — LangfuseClient._paginate walks all pages automatically, so callers just get an iterator of items.

Full reference: https://api.reference.langfuse.com

Gotchas

• fromTimestamp/toTimestamp need full ISO-8601 datetime, not a bare date. The API 400s on 2026-07-01 — LangfuseClient._ts auto-expands a plain YYYY-MM-DD to T00:00:00Z, so passing dates on the CLI is fine; only matters if you're calling the client's methods directly with something else shaped.
• Region-locked keys. Cloud keys are tied to one region (EU cloud.langfuse.com vs US us.cloud.langfuse.com). Wrong host -> 401, not a clearer error. Run health kind first when setting up a new .env — though note health doesn't check auth, only connectivity/version, so a 200 there doesn't mean the keys are valid.
• Unscoped list calls can hang or time out server-side, not just return a lot of data — /sessions and /traces with no date range measured 30s+ per page on a live project in testing. LangfuseClient retries timeouts/connection errors with backoff (default 60s per-request timeout, 5 attempts) and --max-items caps total items, but the real fix is scoping --from/--to tight — don't rely on the cap to make a lazy unscoped query fast, it only stops it from running forever.
• Rate limits. Public API is rate-limited per plan; LangfuseClient retries on 429 honoring Retry-After, and backs off on 5xx. Don't lower max_retries to 0 for bulk pulls. /metrics/daily tends to have tighter limits than /traces. https://langfuse.com/faq/all/api-limits
• GENERATION observations carry the actual prompt/completion sent to the model (input/output fields) — that's what you want for eval datasets or debugging a bad reply, not the trace object itself (which is closer to a request-level summary). promptName/promptVersion on the observation are only populated if the app fetched the prompt via langfuse.get_prompt(...) and passed it through — a hardcoded prompt string in the code won't show a linked version at all.
• Sessions ≠ traces. A session groups multiple traces (e.g. one per turn in a conversation). Pulling session gives you the full multi-turn context in one call; debug-session goes one step further and flattens every generation across all of the session's traces — the shortcut for "what did we send this agent, turn by turn."
• Prompt versioning. label="production" gets whatever is currently deployed; a numeric version pins to an exact snapshot. Use the label when auditing "what's live right now", the version when reproducing a specific past run. To see if a prompt changed recently: pull debug-session for a session before/after the suspected change and diff prompt_version across generations of the same trace_name.
• Large pulls. No built-in checkpointing in extract.py (unlike the Intercom extractor) — for multi-day/multi-GB pulls, chunk by date range in a loop and write incrementally, or extend langfuse_client.py's _paginate with a resume cursor.

Reference Skills

• Instrumenting an app to send traces → skill: langfuse-observability
• Building eval datasets from extracted traces → skill: rag-patterns, llm-engineering