Langfuse Trace Analyzer

Analyze Langfuse trace exports without loading entire files into context. Use jq for targeted extraction.

Trace Structure

{
  "trace": {
    "id", "name", "input", "output",
    "latency", "environment", "tags",
    "sessionId", "userId", "metadata"
  },
  "observations": [
    {
      "id", "name", "type",           // GENERATION | TOOL | CHAIN | SPAN | EVENT
      "parentObservationId",          // hierarchy link
      "model", "modelParameters",     // for GENERATION
      "input", "output",
      "startTime", "endTime", "latency",
      "usageDetails": { "input", "output", "total", "cache_read_input_tokens", "cache_creation_input_tokens" },
      "costDetails": { "input", "output", "total" },
      "totalCost",
      "toolCalls", "toolCallNames"    // tools invoked by this generation
    }
  ]
}

Workflow

Step 1: Normalize the trace file

Langfuse exports contain triple-encoded JSON strings. Normalize first:

bash scripts/normalize.sh /path/to/trace.json

Output: /tmp/langfuse-trace-normalized.json

All subsequent jq queries use this normalized file.

Step 2: Get overview

jq '{
  trace_name: .trace.name,
  total_latency: .trace.latency,
  observations_count: (.observations | length),
  total_cost: ([.observations[] | .totalCost // 0] | add),
  models_used: ([.observations[] | select(.model) | .model] | unique)
}' /tmp/langfuse-trace-normalized.json

Step 3: Targeted analysis

Based on what user wants to investigate, use appropriate jq queries from references/jq-recipes.md.

Common analysis paths:

• Cost analysis: generations with usage, most expensive calls
• Performance: slowest observations, timeline
• Debugging: specific observation input/output, tool calls, errors
• Structure: hierarchy tree, observation types

Key Observation Types

Type	Purpose
GENERATION	LLM call with tokens/cost
TOOL	Tool invocation
CHAIN	Workflow step (LangChain)
SPAN	Generic timed operation
EVENT	Discrete occurrence

Important Notes

Latency field may be inaccurate

The .latency field can be misleading. For accurate duration, calculate from timestamps:

jq '.observations[] | {name, duration_sec: ((.endTime | fromdateiso8601?) - (.startTime | fromdateiso8601?) // .latency)}'

LangChain traces show post-injection data

When using LangChain with dependency injection (e.g., InjectedToolArg), tool call arguments in traces contain injected values, not what the LLM actually sent. This is because Langfuse callbacks capture data after injection happens.

Example: If you see state in tool arguments but your schema doesn't include it - that's injection, not LLM output.

Tips

• Use jq -r for raw string output (no quotes)
• Use .id[0:8] for readable ID prefixes
• Filter with select(.type == "GENERATION") before extracting fields
• Sort with sort_by(-.latency) for descending order

For complete jq recipes, see references/jq-recipes.md.