Skip to main content

Enrichment Agent

A showcase of every waxell-observe enrichment feature: @observe for agent scoping, @step for pipeline stages, score() with all data types, tag() for categorization, and metadata() for structured context. Two child agents handle LLM analysis and quality scoring.

Environment variables

This example requires OPENAI_API_KEY, WAXELL_API_KEY, and WAXELL_API_URL. Use --dry-run to skip real API calls.

Architecture​

Key Code​

Score Data Types​

The evaluator demonstrates all score() data types: numeric, boolean, and categorical.

@waxell.observe(agent_name="enrichment-evaluator", workflow_name="enrichment-scoring")
async def run_enrichment_evaluation(analysis_tokens: int, assessment_tokens: int, waxell_ctx=None):
    waxell.tag("agent_role", "evaluator")

    waxell.score("quality", 0.92)
    waxell.score("relevance", 0.85)
    waxell.score("helpfulness", 0.88)
    waxell.score("safety", True, data_type="boolean")
    waxell.score("category", "informational", data_type="categorical")
    waxell.score("confidence", 0.78, comment="Based on source availability")

Tags and Metadata on the Orchestrator​

The parent agent attaches categorization tags and structured metadata.

@waxell.observe(agent_name="enrichment-orchestrator", workflow_name="enrichment-showcase")
async def run_enrichment_pipeline(query: str, dry_run: bool = False, waxell_ctx=None):
    waxell.tag("demo", "enrichment")
    waxell.tag("intent", "question")
    waxell.tag("domain", "technology")
    waxell.tag("priority", "high")

    waxell.metadata("user_tier", "premium")
    waxell.metadata("request_source", "api")
    waxell.metadata("model_config", {"temperature": 0.7, "max_tokens": 500})
    waxell.metadata("pipeline_version", "2.1.0")

What this demonstrates​

  • All enrichment primitives -- tag(), metadata(), and score() with numeric, boolean, and categorical data types.
  • Multi-agent enrichment flow -- orchestrator sets context, runner does LLM analysis, evaluator records quality scores.
  • @step for pipeline stages -- analyze_input, record_scores, and final_enrichment each recorded as discrete spans.
  • Auto-instrumented LLM calls -- two OpenAI calls (analysis + assessment) captured without extra code.
  • Structured metadata -- nested objects like model_config serialized via _safe_attr.

Run it​

# Dry-run mode (no API key needed)
cd dev/waxell-dev
python -m app.demos.enrichment_agent --dry-run

# Live mode
export OPENAI_API_KEY="sk-..."
python -m app.demos.enrichment_agent

Source​

dev/waxell-dev/app/demos/enrichment_agent.py