feat(annotation): persist n_retrieved_chunks on retrieval records#246
Merged
henrycgbaker merged 3 commits intoJul 6, 2026
Merged
Conversation
github-actions
Bot
added
feature
This was referenced Jun 1, 2026
This was referenced Jul 1, 2026
K (= len(pair.chunks)) is now stamped as record metadata at import, declared on the Argilla retrieval task, modelled on RetrievalAnnotation, and read back by the export fetcher. Mirrors the existing chunk_rank field end-to-end so retrieval metrics can compute true top-K and detect incomplete panels.
henrycgbaker
force-pushed
the
feat/annotation-retrieval-completeness/01-persist-k
branch
from
b232f54 to
30b5118
Compare
ddimmery
approved these changes
Jul 3, 2026
ddimmery
left a comment
ddimmery
left a comment
Collaborator
There was a problem hiding this comment.
This is already live, right? Otherwise there would need to be some kind of migration to deal with adding columns mid-run.
| chunk_id=metadata.get("chunk_id", ""), | ||
| doc_id=metadata.get("doc_id", ""), | ||
| chunk_rank=metadata.get("chunk_rank", 0), | ||
| n_retrieved_chunks=metadata.get("n_retrieved_chunks", 0), |
Collaborator
There was a problem hiding this comment.
This zero is a sentinel, right? I ask because I assume downstream we're assuming k > 0 (e.g. in the task definition) and then use it as a denominator in metrics, so we'll need to watch it there.
Collaborator
Author
There was a problem hiding this comment.
yes, 0 is sentinel - real values guaranteed ≥1 by Argilla schema (IntegerMetadataProperty(..., min=1)). I could change to None, only (non-major) issue is that this is close in the pipeline to csv/df export - so this would be adding non-int datatype which i think pandas coerces into NaN. In reality this is a non-issue as it's not reachable/edge case due to IntegerMetadataProperty
Collaborator
There was a problem hiding this comment.
maybe -1 to keep it an integer but make it very clear it's different?
Collaborator
Author
yes this is already live (I had a non-git tracked backfill script) |
Addresses review feedback on #246: 0 could theoretically be confused with a real count; -1 falls outside the valid domain (real K is always >=1) and keeps the field an int rather than switching to None.
henrycgbaker
deleted the
feat/annotation-retrieval-completeness/01-persist-k
branch
henrycgbaker
added a commit
that referenced
this pull request
Jul 6, 2026
**Stack (6 PRs):** #246 (merged) → #247 → #268 → #248 → #269 → #249 **Chain goal:** Persist K (number of retrieved chunks per query) on retrieval records, surface it as panel-completeness columns + sidecar at export time, ship a live `annotation status` CLI on top, and add an opt-in `--tag-partial-panels` advisory write so annotators can filter partially-done panels in the Argilla UI. **This PR (2/6):** The standalone, export-agnostic panel-completeness calculator, plus the shared retrieval-walk primitive it and #268 both consume. Split out of what was originally a single ~1000-line PR; wiring completeness into the retrieval export path ships separately in #268. --- ## Scope ### Completeness aggregator - New `core/annotation/completeness.py`: groups retrieval snapshots by `record_uuid`, distinct-by-`chunk_id`, derives per-panel facts, emits `CompletenessReport` (`by_uuid` + `summary`). Aggregation is split into a typed `_PanelAccumulator` plus `_aggregate_snapshots` (grouping) / `_summarize_groups` (transform) passes. ### Shared retrieval walk - `core/annotation/export_fetcher.py`: `RetrievalRecordSnapshot` + `walk_retrieval_records` — one Argilla scroll, no status filter (so records lacking terminal responses are still seen for the integrity check), feeding both this module's completeness pass and #268's row emission. Plus `resolve_task_purposes`, a topology lookup extracted from `fetch_task` (no behaviour change). ### Schema - `core/schemas/annotation_export.py`: `CompletenessSummary` + `KBucketStat` models for the export sidecar (populated by #268). ### Tests - New `test_completeness.py`: happy path, distinct-by-chunk_id, orphan exclusion, integrity warnings (records-vs-K + mixed-backfill within panel), all-unknown-K warning, K-bucket cross-tab + per-K histogram, STRICT `panel_complete` (discards do NOT count), `n_discarded_chunks` subset of `n_annotated_chunks`, calibration walked when topology declares it. ## Why STRICT `panel_complete` The computed `panel_complete: bool` is STRICT: True iff every K chunk in the panel has at least one **submitted** response. Discarded responses are abstentions, not judgements (pragmata's `DiscardReason` enum is refusal-only: `INVALID_OR_UNREALISTIC` / `UNCLEAR` / `OUTSIDE_REVIEWER_EXPERTISE`), so treating them as covered would feed unjudged chunks into NDCG@K / precision@K denominators as 0-relevance labels. The bool is derivable from the count columns, so a consumer wanting the permissive policy computes `n_annotated_chunks == n_retrieved_chunks` in one line. The retriever is threshold-based, so dropping partial panels biases metrics toward small K (MNAR); this module is the shared calculation both the export path (#268) and the live status path (#248) build on — one aggregator, two consumers. ## Test plan - [x] `uv run python -m pytest tests/unit` (1040 passing)
henrycgbaker
added a commit
that referenced
this pull request
Jul 7, 2026
…ry) (#268) **Stack (6 PRs):** #246 → #247 → #268 → #248 → #269 → #249 **This PR (3/6):** Split out of #247 (was 1001 lines / XL). Wires the panel-completeness calculator (#247) into the retrieval export path. Stacked on #247. --- ## Scope ### Schema - `core/schemas/annotation_export.py`: - `RetrievalExportRow` += `panel_complete: bool`, `n_annotated_chunks: int`, `n_submitted_chunks: int`, `n_discarded_chunks: int`, `n_records_seen: int`. All defaulted; appended at the tail so existing column order is preserved. - New `KBucketStat` (per-K-bucket panel counts) and `CompletenessSummary` (aggregates + `by_k_bucket` cross-tab + full per-K histogram) Pydantic models. - `AnnotationExportMeta` += `completeness_summary: CompletenessSummary | None` and `completeness_status: Literal["ok","failed","not_requested"]`. ### Single-walk retrieval pipeline - `core/annotation/export_fetcher.py`: new `walk_retrieval_records` (one Argilla scroll per dataset, no status filter — sees discards-only records too so the integrity check has the full record set) + `RetrievalRecordSnapshot` dataclass + `fetch_retrieval_from_records` (status filter applied in Python at typed-row construction). - `core/annotation/export_runner.py`: when `Task.RETRIEVAL in tasks`, walks once and feeds both the typed-row projection and the completeness aggregator. `compute_completeness` failures degrade the sidecar (sets `completeness_status="failed"`) without losing the already-fetched export. ### API - `annotation/__init__.py`: lazy re-exports for `CompletenessSummary`, `KBucketStat`. ### Tests - `test_export_runner.py`: `write_export_csv` stamps completeness columns by `record_uuid`; sidecar carries `completeness_summary`; `completeness_status` ok/failed/not_requested set explicitly. - `test_annotation_export.py`: `RetrievalExportRow` field-order locked. ## Why The retriever is threshold-based, so dropping partial panels biases the metrics toward small K (MNAR). Surfacing both per-row predicate columns AND a bucketed sidecar aggregate lets the eval pipeline either filter complete panels per-K or apply condensed-list scoring without re-querying the data layer. The single-walk shape (one retrieval scroll feeds both fetch_task and compute_completeness) costs one extra in-memory pass over already-fetched records but spares the Argilla server a second full scroll per export — meaningful on the in-flight batches (~hundreds of panels). The STRICT `panel_complete` policy question (which shape to ship — this PR assumes (A) from #247) is discussed on #247, since that's where the predicate is actually computed; this PR just surfaces it as a column. ## Test plan - [ ] `uv run python -m pytest tests/unit/core/annotation/test_export_runner.py tests/unit/core/schemas/test_annotation_export.py`
henrycgbaker
added a commit
that referenced
this pull request
Jul 7, 2026
…249) **Stack (6 PRs):** #246 → #247 → #268 → #248 → #269 → #249 **This PR (6/6):** Adds the advisory `needs_completion` tag write onto `annotation status`. - `status --tag-partial-panels`: stamps `needs_completion` on the **unresolved chunks of PARTIAL panels** (`0 < submitted < K`) and clears stale tags, sharing the single status walk. Annotators filter `needs_completion=true` in the Argilla UI to focus on the records that will complete a partially-done panel. - **Partial-only gate** (not "any incomplete panel"): a fully-unstarted panel is never tagged, so the tag stays a selective mop-up signal instead of ≈ everything pending. - **Multi-domain + prod/cal-split aware**: a panel whose chunks are split across a workspace's production + calibration datasets (per-item calibration) is evaluated as one unit; writes route to each chunk's owning dataset. Reuses `metadata_ops` for the replace-safe upsert. Stacked on #269. Also re-exports `TagResult` from the `pragmata.annotation` facade: `StatusReport.tag_result` is a public field of this type, but the facade export was pre-declared two PRs ago (before `panel_status.py` existed), dropped once that PR's version of the module still didn't define the class, and never re-added once this PR added it for real.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack (6 PRs): #246 → #247 → #268 → #248 → #269 → #249
Chain goal: Persist K (number of retrieved chunks per query) on retrieval records, surface it as panel-completeness columns + sidecar at export time, ship a live
annotation statusCLI on top, and add an opt-in--tag-partial-panelsadvisory write so annotators can filter partially-done panels in the Argilla UI.This PR (1/6): Foundation. Stamps
n_retrieved_chunkson retrieval record metadata at import and surfaces it onRetrievalAnnotation. Ships nothing user-visible on its own — the export columns (#247/#268) and live status CLI (#248/#269) consume it.Scope
Record building
core/annotation/record_builder.py: add"n_retrieved_chunks": len(pair.chunks)to every chunk-record's metadata dict.Task definition
core/annotation/argilla_task_definitions.py: declareIntegerMetadataProperty("n_retrieved_chunks", min=1, visible_for_annotators=False)on retrieval task settings.Schema
core/schemas/annotation_export.py:RetrievalAnnotation.n_retrieved_chunks: intfield (mirrors the existingchunk_rank: int).Export readback
core/annotation/export_fetcher.py:_build_rowretrieval branch readsmetadata.get("n_retrieved_chunks", 0)(defensive default for records imported before this PR).Tests
test_n_retrieved_chunks_stamped_on_every_chunk_recordintest_import.pylocks the persist contract.test_argilla_task_definitions.pyupdated.test_annotation_export.py,test_export_runner.py,test_export_constraint_checks.py,test_iaa_runner.py,test_export_api.py.Why
K varies per query (~15% K<5, 61% K=5, 24% K>5; the retriever is threshold-based, not top-K), so the export side can't infer K from
chunk_rankor from record-counting alone — it needs a persisted SSOT to drive true top-K metrics in the consumer pipeline (precision@K, recall@K, NDCG@K, MRR@K).End-to-end mirror of the existing
chunk_rankfield: stamped at import, declared on the task, modelled onRetrievalAnnotation, read back by the fetcher.A one-off backfill script for records imported before this PR (in-flight batches) lives outside the supported CLI/API surface — retrospective ops migration, not packaged functionality.
Base
feat/querygen-resumable-stack— kitchen-sink integration branch carrying the full dep surface this stack assumes (Locale,atomic_write_text,LogicalConstraint,WIDGET_FIELD_PLACEHOLDERS, theexport_constraint_checksrename). Rebases cleanly tomainonce those upstream stacks land.Test plan
uv run python -m pytest tests/unit(863 passing)