Converge getEmbeddingInput()'s obs_group_concept_name read onto getObsGroupConceptName()

[dkayiwa]

Summary

QueryDocument.getEmbeddingInput() reads the obs-group concept name inline from the metadata map, while #60 added a centralizing accessor for the same field. They use different normalization, so the read the #60 accessor was meant to centralize still has a second, divergent implementation inside the model itself. Converge them — but it is a deliberate decision, not a trivial refactor, because converging changes embedding behavior (see below).

Current state

• getObsGroupConceptName() (added in Add typed getObsGroupUuid() / getObsGroupConceptName() accessors on QueryDocument #60) → metadataString(FIELD_OBS_GROUP_CONCEPT_NAME): trim + blank→null ("clean String", identity semantics).
• getEmbeddingInput() (QueryDocument.java, the group-prefix branch) reads the field directly:

  Object groupName = metadata.get(FIELD_OBS_GROUP_CONCEPT_NAME);
  if (groupName instanceof String && !((String) groupName).isEmpty()) {
      sb.append(groupName).append(" — ");   // raw value, no trim; skip-empty (not skip-blank)
  }

  i.e. raw value + skip-empty, no trim.

Failure mode this prevents (silent divergence)

Low probability, but the same drift class #60 addressed — only internal: if querystore later changes how obs_group_concept_name is normalized, the canonical getObsGroupConceptName() accessor updates in one place, but getEmbeddingInput()'s inline raw read does not, so the consumer-facing value and the embedding-prefix could drift for the same field.

Why it was deferred (the real decision)

This is not a free refactor — getEmbeddingInput() feeds the embedding vectors, so changing its read changes the vectors:

• Option A — make getEmbeddingInput() use the accessor (trim + blank→null). Cleaner, single read shape, and arguably more correct (a whitespace-only group name currently emits a " — " prefix; trimming would drop it). But it changes the embedding-input text for any record whose group name has surrounding/!empty-but-blank whitespace → different vectors → requires re-running the retrieval eval rubric (RetrievalQualityEvalTest + the MySQL/ES integration variants) to confirm no regression. In practice group names come from ConceptNameUtil.getPreferredName and are rarely padded, so the real-data impact is likely nil — but it must be verified, not assumed.
• Option B — give the accessor getEmbeddingInput()'s raw/no-trim semantics. Avoids any embedding change and lets getEmbeddingInput() reuse it cleanly, but contradicts Add typed getObsGroupUuid() / getObsGroupConceptName() accessors on QueryDocument #60's "clean String" intent for the consumer-facing accessor (consumers would get untrimmed values).

A decision between A and B is needed before converging.

Scope / acceptance

• Decide A vs B (recommend A — "clean String" everywhere — if the eval confirms no regression).
• Converge getEmbeddingInput()'s group-name read onto the chosen shape so the field is read one way in the model.
• If Option A: re-run the eval rubric and record that recall@30 is unchanged (still ≥ the 0.40 gate / ~0.888 baseline) in the PR.

Notes

• Not a correctness bug. getEmbeddingInput() is correct and tested today; this is a reuse/consistency tidy, like Add typed getObsGroupUuid() / getObsGroupConceptName() accessors on QueryDocument #60 itself.
• Surfaced as the deferred Phase-1 trace-outward item during Add typed getObsGroupUuid() / getObsGroupConceptName() accessors on QueryDocument #60's hardening.

[Dharshika-11]

Hi, I'd like to work on this issue.

[dkayiwa]

@Dharshika-11 oh yes go ahead! 👍