fix(knowledge): walk a large bounded set on the row before ranking it exactly by waleedlatif1 · Pull Request #8106 · simstudioai/sim

waleedlatif1 · 2026-09-21T17:00:24Z

Summary

A member who can read most of a large source, searching with that source selected as a filter, resolved a bounded permitted set of tens of thousands of documents and then ranked every chunk of it exactly on both legs. The vector leg read every chunk's projected vector; the keyword leg materialized every chunk of the set before matching the term, so its cost followed the set's size rather than the term's matches. On a cold cache each leg outran its budget and the search returned nothing
A bounded set past PERMITTED_EXACT_DOCUMENT_LIMIT (5,000 documents) is now walked on the row first: the plan's source and ACL decide readability on the rows the walk visits and the walk stops at its tuple cap. If the walk cannot fill its pool the exact ranking that was always complete takes over, so recall is never below today's and the usual cost is the walk's. Measured on a cold cache, the walk restricted to the source costs milliseconds where the exact ranking costs over a second
The keyword leg treats the same set as a narrow on-row reader: Tin windows where Tin serves, otherwise the unfenced GIN shape whose cost follows what the term matches. Deployments without Tin keep GIN throughout
Sets under the limit, unbounded readers and broad readers keep their existing paths unchanged; the on-row walk is hoisted into one helper and used by both callers
The size limit sits where exact ranking reads about as many rows as the walk's tuple cap

Type of Change

Bug fix

Testing

Six new tests: a large bounded set walks on the row and sends no identifiers, an underfilled walk falls back to the exact ranking with the set, a set under the limit ranks exactly without a walk; the keyword leg ranks the large set with Tin as a narrow reader, falls back to an unfenced GIN read, and keeps the fenced read under the limit. Each was verified to fail with its half of the fix reverted
vitest over lib/knowledge, app/api/knowledge, app/api/v1/knowledge, app/api/v2/knowledge, lib/copilot/tools/server/knowledge: 218 files, 3548 tests passing
bun run lint, bun run check:audits, docs-manifest:check and native tsc --noEmit pass

Checklist

Code follows project style guidelines
Self-reviewed my changes
Tests added/updated and passing
No new warnings introduced
I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

… exactly A member reading most of a large source, with that source selected as a filter, enumerated a bounded permitted set of tens of thousands of documents and then ranked every chunk of it exactly on both legs: the vector leg read every chunk's projected vector, and the keyword leg materialized every chunk of the set before it matched the term. Cold, each leg outran its budget and the search returned nothing. A bounded set past a size limit is now walked on the row first, where the plan's source and ACL decide readability and the walk stops at its tuple cap, and ranked exactly only when the walk cannot fill its pool, so recall is never below the exact ranking's. The keyword leg treats the same set as a narrow on-row reader: Tin windows where Tin serves, otherwise the GIN shape whose cost follows the term's matches. Sets under the limit keep their exact paths.

vercel · 2026-09-21T17:00:30Z

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
docs	Skipped		Sep 21, 2026 5:29pm UTC

cubic-dev-ai

All reported issues were addressed across 2 files

_{Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic}

greptile-apps · 2026-09-21T17:04:28Z

Confidence Score: 5/5

The PR appears safe to merge; the previously reported vector refill and Tin-to-GIN offset issues are fixed, and no new actionable defects remain.

Summary

This PR optimizes searches over large bounded permitted-document sets while preserving exact-ranking fallbacks.

Vector search first performs a bounded on-row graph walk, then falls back to exact ranking when the walk or subsequent hydration cannot supply enough authorized candidates.
Exact refills exclude candidates already read so each refill supplies a fresh ranking window.
Keyword search treats large bounded sets as on-row readers and uses Tin when available, with a first-page GIN fallback that avoids carrying offsets between incompatible rankers.
Tests cover the optimized paths, exact fallbacks, refill behavior, size threshold, and later-page Tin behavior.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Search over permitted documents] --> B{Bounded set at least 5,000 documents?}
    B -->|No| C[Use existing exact or bounded path]
    B -->|Yes, vector| D[Walk graph with on-row visibility]
    D --> E{Pool filled and hydration sufficient?}
    E -->|Yes| F[Return hydrated results]
    E -->|No| G[Exact-rank permitted set]
    G --> H[Exclude previously read chunks on refill]
    H --> F
    B -->|Yes, keyword| I{Tin available?}
    I -->|Yes| J[Rank through Tin windows]
    J --> K{First page still short?}
    K -->|Yes| L[Restart with unfenced GIN ranking]
    K -->|No or later page| M[Return Tin page]
    I -->|No| L

_{Reviews (5) · Last reviewed commit: "fix(knowledge): hand only a large bounde..."}

…ing once hydration runs it short The walk decides readability on the projection row, which is broader than the document predicate hydration applies, so a pool the walk filled can still run short of readable rows. The refill for a large bounded set is now the exact ranking, complete over the set, placed behind the rows already read so the pages keep their offsets.

waleedlatif1 · 2026-09-21T17:08:26Z

@greptile

waleedlatif1 · 2026-09-21T17:08:27Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-09-21T17:08:37Z

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

…dy read The refill's exact ranking excludes the chunks the pool already holds inside the statement, so every refill is a full window of fresh rows rather than a window thinned by the rows the walk found first.

cubic-dev-ai

All reported issues were addressed across 2 files

_{Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic}

waleedlatif1 · 2026-09-21T17:15:43Z

@greptile

waleedlatif1 · 2026-09-21T17:15:44Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-09-21T17:15:51Z

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

…he GIN ranking A narrow reader's page is left short by design once the widest window cannot fill it; a large bounded set's read was exhaustive before, so its widest window that still falls short now hands the page to the GIN ranking, which covers every match.

waleedlatif1 · 2026-09-21T17:21:21Z

@greptile

waleedlatif1 · 2026-09-21T17:21:23Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-09-21T17:21:30Z

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

cubic-dev-ai

No issues found across 2 files

Confidence score: 5/5

Automated review surfaced no issues in the provided summaries.
No files require special attention.

_{Re-trigger cubic}

… ranking Tin and GIN order candidates differently, so an offset advanced through one ranking cannot resume the other. A large bounded set's first page that Tin's widest window cannot fill goes to GIN; a later page stays with Tin and is left short as a narrow reader's is.

waleedlatif1 · 2026-09-21T17:29:56Z

@greptile

waleedlatif1 · 2026-09-21T17:29:57Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-09-21T17:30:04Z

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

cubic-dev-ai

No issues found across 2 files

Confidence score: 5/5

Automated review surfaced no issues in the provided summaries.
No files require special attention.

_{Re-trigger cubic}

cubic-dev-ai Bot reviewed Sep 21, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/search/queries.ts Outdated

greptile-apps Bot reviewed Sep 21, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/search/queries.ts Outdated

vercel Bot temporarily deployed to Preview September 21, 2026 17:08 Inactive

greptile-apps Bot reviewed Sep 21, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/search/queries.ts

fix(knowledge): rank a large bounded set's refill past the rows alrea…

fd65d98

…dy read The refill's exact ranking excludes the chunks the pool already holds inside the statement, so every refill is a full window of fresh rows rather than a window thinned by the rows the walk found first.

cubic-dev-ai Bot reviewed Sep 21, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/search/queries.ts

vercel Bot temporarily deployed to Preview September 21, 2026 17:15 Inactive

vercel Bot temporarily deployed to Preview September 21, 2026 17:21 Inactive

cubic-dev-ai Bot reviewed Sep 21, 2026

View reviewed changes

greptile-apps Bot reviewed Sep 21, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/search/queries.ts Outdated

vercel Bot temporarily deployed to Preview September 21, 2026 17:29 Inactive

cubic-dev-ai Bot reviewed Sep 21, 2026

View reviewed changes

waleedlatif1 merged commit c9d73bb into staging Sep 21, 2026
25 checks passed

waleedlatif1 deleted the fix/source-filter-reach branch September 21, 2026 17:33

This was referenced Sep 21, 2026

v0.8.49: knowledge search improvements, tables perf improvements #8108

Merged

fix(knowledge): keep the GIN ranking for the rest of a search once a page was handed to it #8110

Merged

Conversation

waleedlatif1 commented Sep 21, 2026

Summary

Type of Change

Testing

Checklist

Uh oh!

vercel Bot commented Sep 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

greptile-apps Bot commented Sep 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 5/5

Summary

Uh oh!

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

cubic-dev-ai Bot commented Sep 21, 2026

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

cubic-dev-ai Bot commented Sep 21, 2026

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

cubic-dev-ai Bot commented Sep 21, 2026

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

waleedlatif1 commented Sep 21, 2026

Uh oh!

cubic-dev-ai Bot commented Sep 21, 2026

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

This branch was previously deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

vercel Bot commented Sep 21, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

greptile-apps Bot commented Sep 21, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading