Vector SNOMED grounding mapping: hybrid retrieval update + live T22 gold validation

[phiwi]

Summary

This issue tracks the vector-SNOMED grounding mapping update for Table 22 gold evaluation: why we changed it, what was implemented, and which validation runs are currently active.

Idea / motivation

• Baseline lexical grounding is robust but can miss semantically close SNOMED candidates when wording diverges.
• We added vector-assisted retrieval to improve candidate recall while preserving lexical precision through reranking.
• Goal: compare vector vs non-vector grounding directly against gold SNOMED IDs from Table 22, without concept/rule extraction confounds.

Implemented change

1) Grounding-only eval path (gold SNOMED comparison)

• Added dedicated evaluator:
  • src/cardio_graph_core/evaluation/table22_snomed_grounding_only_eval.py
• Reads gold SNOMED concepts from table_22_manual_snomed.json and measures exact-ID hit accuracy.
• Supports both modes:
  • --mode non-vector
  • --mode vector
• Writes per-item predictions and aggregate metrics JSON.

2) Separate Slurm jobs for clean A/B comparison

• Non-vector launcher:
  • slurm/run_table22_snomed_grounding_only_nonvector.sh
• Vector launcher:
  • slurm/run_table22_snomed_grounding_only_vector.sh
• Both run on g3:11433 (model and embedding endpoints aligned to avoid prior endpoint drift).

3) Vector retriever warning fix

• Updated vector retriever to avoid Neo4j warning-prone direct property access in Cypher:
  • src/cardio_graph_core/extraction/vector_candidate_retriever.py
• Query now returns node, score; concept-id/term fallbacks resolved in Python.
• This removes repeated UnknownPropertyKeyWarning noise and keeps fallback behavior explicit.

Currently running tests (live)

• 625300 — t22_gold_vec — RUNNING
• 625298 — t22_gold_nonvec — RUNNING

Current log signals:

• Vector log (slurm/run_table22_snomed_grounding_only_vector.log)
  • vector grounding enabled against snomed_term_embeddings_4096
  • connected to SNOMED DB
  • ongoing candidate truncation messages (expected while iterating terms)
• Non-vector log (slurm/run_table22_snomed_grounding_only_nonvector_625298.log)
  • connected to SNOMED DB
  • ongoing candidate truncation messages (expected)

Expected output artifacts on completion:

• Vector:
  • docs/table22_snomed_grounding_compare/grounding_only/vector_job_625300/vector_eval.json
• Non-vector:
  • docs/table22_snomed_grounding_compare/grounding_only/nonvector_job_625298/nonvector_eval.json

Next update

I will post final accuracy deltas (vector minus non-vector) and improved/regressed examples once both runs finish.