Refined Stage 3 PromptFlow atom extraction pipeline

• Day: 2026-05-25
• Time: 11:20 to 11:30
• Project: Dev
• Workspace: WP 2: Operational
• Status: In Progress
• Priority: HIGH
• Assignee: Matías Nehuen Iglesias
• Tags: Promptflow, Atom-Extraction, Debugging, Schema, Workflow, Validation

Description

Session Goal

Stabilize and improve the Stage 3 atom extraction workflow by diagnosing PromptFlow failures, tightening extraction caps, and deciding whether to keep the native LLM node or revert to the proven wrapper-based architecture.

Key Activities

• Reviewed the deterministic Stage 3 extraction-plan builder as the correct seam between router output and downstream extraction.
• Diagnosed multiple PromptFlow failures and isolated them to configuration and wiring issues rather than atom schema logic:
  • connection-name mismatch (openai vs open_ai_connection)
  • missing model parameter in the OpenAI node
  • incorrect placement of model under parameters instead of inputs
  • chat-node prompt syntax issues and a Jinja rendering bug where max_cases was hardcoded instead of templated
• Evaluated the Stage 3 atom mining plan and identified under-extraction in the current spread20 setup, especially for case counts.
• Proposed a more controlled Stage 3 atom mining flow with strict candidate schemas, reporting, and a small mixed smoke-test sample before semantic deduplication.
• Drafted a patch to apply_family_overrides so extraction limits adapt by content family, format, sensitivity, and title signals.
• Considered a fallback migration back to the stable llm_wrapper / legacy function-schema architecture used in Stage 2 if native PromptFlow remained unstable.
• Defined a cautious validation sequence: 5-row smoke test, schema validation, report generation, and manual quality checks before scaling.

Achievements

• Clarified that the main failure mode was PromptFlow configuration drift, not the extraction schema itself.
• Established that the wrapper-based architecture is the stable reference implementation for Stage 3.
• Identified concrete fixes for routing, cap handling, prompt templating, and model wiring.
• Improved the extraction plan by adding traceability and override-awareness for sensitive/personal narrative routing.

Pending Tasks

• Apply the PromptFlow DAG and prompt fixes, then rerun the Stage 3 smoke test.
• Verify that total_max_cases increases correctly after the family-specific override patch.
• Add or restore a policy-level validator for cap enforcement.
• Decide whether to keep the native PromptFlow node or fully migrate Stage 3 to the Python wrapper pattern.
• Run manual quality checks on the smoke-test output before any broader dataset execution.

Evidence

• source_file=2026-05-25.sessions.jsonl, line_number=0, event_count=0, session_id=20d194467206e195ae967f90906d80483553fea08654b0756b72cf538a7463de
• event_ids: []