@absolutejs/rag 0.22.15 → 0.24.0

package/CHANGELOG.md

@@ -6,6 +6,18 @@ This file is generated by `absolute-changelog` from the entries in
 `changelog/`. Edit an entry, not this file — and add new ones under
 `changelog/unreleased/`.
 
+## 0.24.0 — 2026-11-03
+
+### Added
+
+- **Bounded review batches with durable unreviewed drafts, fenced host checkpoints, explicit resume and preserved partial verdicts.** (`createResearch`, `ResearchConfig`, `ResearchInput`, `ResearchResult`, `ResearchReviewCheckpoint`)
+
+## 0.23.0 — 2026-11-02
+
+### Added
+
+- **Opt-in bounded evidence correction with fresh independent review, reserved retrieval, retained review history and atomic collection fact coverage.** (`createResearch`, `ResearchConfig`, `ResearchResult`, `ResearchTask`)
+
 ## 0.22.15 — 2026-11-01
 
 ### Fixed

package/RESEARCH.md

@@ -305,3 +305,44 @@ to a different model. A stronger reviewer does not repair missing retrieval.
 An invalid planner schema no longer discards retrieved evidence: extraction and
 review continue without retry, while the result remains partial and the failed
 plan retains unknown settlement. Budget denial and cancellation still stop work.
+
+## Correcting reviewed evidence gaps
+
+Set `evidenceRepairs: 1` (maximum 2, default 0) to allow a bounded correction after incomplete collection coverage or rejected fields. `repairSearches` and `repairReads` reserve part of the existing total retrieval limits for review-identified gaps; they do not increase those limits. Each correction extracts again and independently reviews every new leaf. Drafts and review reasons are hints, never evidence. `reviewPasses` retains the reviewed drafts and coverage for audit. Provider failures and admission denial stop work without automatic replay.
+
+Collection tasks can set `minimumSupportedFields: { qualificationFacts: 1 }` alongside `requiredFields: ["name", "website"]` to require a minimum number of supported string leaves in an atomic fact array. Unsupported leaves remain rejected; callers must project only supported fields when `data` is null.
+
+Budget for each correction's optional retrieval plan, extraction and review, including configured schema repairs. Every model and retrieval operation follows the same admission and settlement path as initial research. Correction is disabled unless explicitly configured.
+
+## Durable batched review (0.24)
+
+Review now runs in small field batches rather than requiring the complete review
+in one model response. `reviewOutputTokens` defaults to the smaller of the global
+output limit and 2,048; `reviewBatchSize` defaults to 12 and is further bounded to
+one field per 256 output tokens (eight fields at the default review ceiling).
+Price every batch, validation repair and evidence-correction pass separately.
+`limits.fields` bounds work per review attempt; excess fields remain in the draft
+for explicit continuation, rather than invalidating the extraction.
+
+`result.draft` is **unreviewed**, never a supported answer. Only `fields` carry
+review verdicts; `data` still requires all fields to pass. `documents` retain
+original text for both search snippets and fetched pages independently of the
+model-facing excerpt allowance.
+
+For durable operation, supply `input.onReviewCheckpoint` and await a tenant-bound,
+compare-and-swap write in that callback. The runtime calls it before extraction,
+before each review batch, and after each completed batch. Persistence failure
+stops execution. Reopen trusted stored state with `input.resumeReview` to reuse
+sources, extraction and completed batches. Task/model/batch identity must match.
+Checkpoints and drafts are private host state, never client-supplied proof.
+
+A `pending` checkpoint cannot resume: the provider outcome or checkpoint write
+requires reconciliation. Admission denial leaves a safe continuation. A host
+using its own metering may implement `isReviewFailureSettled` only when it can
+prove the failed generation completed and its usage is durably accounted for;
+this hook also applies to extraction failures. It never triggers an automatic
+retry. Package-budget failures remain conservatively pending. `finished` means
+review processing finished, not that every claim was supported. Reopening a
+finished checkpoint performs no provider work. Checkpoints during evidence
+correction remain pending until the new extraction is safely retained, preventing
+replay of uncertain correction calls.

package/changelog.json

@@ -2,6 +2,39 @@
 	"contract": 1,
 	"name": "@absolutejs/rag",
 	"releases": [
+		{
+			"changes": [
+				{
+					"kind": "added",
+					"summary": "Bounded review batches with durable unreviewed drafts, fenced host checkpoints, explicit resume and preserved partial verdicts.",
+					"symbols": [
+						"createResearch",
+						"ResearchConfig",
+						"ResearchInput",
+						"ResearchResult",
+						"ResearchReviewCheckpoint"
+					]
+				}
+			],
+			"date": "2026-11-03",
+			"version": "0.24.0"
+		},
+		{
+			"changes": [
+				{
+					"kind": "added",
+					"summary": "Opt-in bounded evidence correction with fresh independent review, reserved retrieval, retained review history and atomic collection fact coverage.",
+					"symbols": [
+						"createResearch",
+						"ResearchConfig",
+						"ResearchResult",
+						"ResearchTask"
+					]
+				}
+			],
+			"date": "2026-11-02",
+			"version": "0.23.0"
+		},
 		{
 			"changes": [
 				{