Bump langchain-community from 0.0.10 to 0.3.27 - #2
Open
dependabot[bot] wants to merge 1 commit into
Open
Conversation
Bumps [langchain-community](https://github.com/langchain-ai/langchain) from 0.0.10 to 0.3.27. - [Release notes](https://github.com/langchain-ai/langchain/releases) - [Commits](langchain-ai/langchain@langchain-robocorp==0.0.10...langchain==0.3.27) --- updated-dependencies: - dependency-name: langchain-community dependency-version: 0.3.27 dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com>
dependabot
Bot
added
dependencies
ringochen06
added a commit
that referenced
this pull request
Jul 28, 2026
Three measured improvements, each isolated so the attribution is clean, plus one hypothesis of mine that the data refuted. Filters now scope retrieval instead of truncating results A latent bug: filters were applied after the top-N candidate cut, so a scoped query searched the whole corpus and then deleted the non-matching results. _allowed_ids restricts the candidate universe first, and framework joins jurisdiction and category as a filter dimension, wired through the facade, the CLI and the API. The hypothesis this was built to test was wrong. Section 4b predicted q11-reidentification failed because 470 GDPR chunks drowned it; scoping to framework=GA4GH changed nothing, still 0 of 4. The real cause was additive RRF. The GA4GH prohibition ranked BM25 #2 and dense #50, and summing rewards corroboration, so six chunks of general privacy prose that were merely respectable in both channels beat it out of the top-8. Enlarging the dense pool did not help either; the sweep is flat from 12/48 to 128/128. Fusion now takes each chunk's best channel rather than the sum, which gains 4 points of recall@8 and 4 of recall@5 and gives up rank-1 accuracy and MRR — the same trade direction as the lexical-weighted pools, for the same reason. REGBOT_FUSION=sum restores classic RRF. Chunking follows headings Chunks break at heading boundaries, splitting only inside sections that exceed the size cap. Straddling chunks fall from 146 of 689 to 15 of 763, so a citation points at one article rather than two unrelated obligations. Recall@1 rises from 0.299 to 0.533. Read precision@8 with care here: re-chunking changes which chunks contain an anchor phrase, so it changes the gold set itself. Several queries went from two gold chunks to one, which mechanically caps precision@8 at 1/8. Recall is not strictly comparable across chunking schemes, and the durable justification for this change is citation quality rather than the metric. Taiwan on official law, and why the other five are not Only Taiwan could be migrated: the Ministry of Justice English text of both Acts, 198 articles. Singapore and Japan return a table of contents rather than the statute — the Singapore file contained zero occurrences of "shall not" and the Japanese one 109 collapsed-section markers — and ingesting a table of contents would pollute retrieval with structural headings that match query vocabulary and contain no rules, so both were discarded. Hong Kong and Korea serve single-page apps. q02 fell from 1.00 to 0.00, and it is the most informative number in the set. Section 3 warned that the P2 queries scoring 1.00 measured the jurisdiction filter rather than retrieval quality. With Taiwan on three summary chunks any retrieval found the answer; with 207 chunks of real statute the operative withdrawal provision sits at rank 14. A hollow measurement was replaced by a real one. Recorded in eval_results.md section 4c. Test count 117 to 122.
ringochen06
added a commit
that referenced
this pull request
Jul 28, 2026
Three measured improvements, each isolated so the attribution is clean, plus one hypothesis of mine that the data refuted. Filters now scope retrieval instead of truncating results A latent bug: filters were applied after the top-N candidate cut, so a scoped query searched the whole corpus and then deleted the non-matching results. _allowed_ids restricts the candidate universe first, and framework joins jurisdiction and category as a filter dimension, wired through the facade, the CLI and the API. The hypothesis this was built to test was wrong. Section 4b predicted q11-reidentification failed because 470 GDPR chunks drowned it; scoping to framework=GA4GH changed nothing, still 0 of 4. The real cause was additive RRF. The GA4GH prohibition ranked BM25 #2 and dense #50, and summing rewards corroboration, so six chunks of general privacy prose that were merely respectable in both channels beat it out of the top-8. Enlarging the dense pool did not help either; the sweep is flat from 12/48 to 128/128. Fusion now takes each chunk's best channel rather than the sum, which gains 4 points of recall@8 and 4 of recall@5 and gives up rank-1 accuracy and MRR — the same trade direction as the lexical-weighted pools, for the same reason. REGBOT_FUSION=sum restores classic RRF. Chunking follows headings Chunks break at heading boundaries, splitting only inside sections that exceed the size cap. Straddling chunks fall from 146 of 689 to 15 of 763, so a citation points at one article rather than two unrelated obligations. Recall@1 rises from 0.299 to 0.533. Read precision@8 with care here: re-chunking changes which chunks contain an anchor phrase, so it changes the gold set itself. Several queries went from two gold chunks to one, which mechanically caps precision@8 at 1/8. Recall is not strictly comparable across chunking schemes, and the durable justification for this change is citation quality rather than the metric. Taiwan on official law, and why the other five are not Only Taiwan could be migrated: the Ministry of Justice English text of both Acts, 198 articles. Singapore and Japan return a table of contents rather than the statute — the Singapore file contained zero occurrences of "shall not" and the Japanese one 109 collapsed-section markers — and ingesting a table of contents would pollute retrieval with structural headings that match query vocabulary and contain no rules, so both were discarded. Hong Kong and Korea serve single-page apps. q02 fell from 1.00 to 0.00, and it is the most informative number in the set. Section 3 warned that the P2 queries scoring 1.00 measured the jurisdiction filter rather than retrieval quality. With Taiwan on three summary chunks any retrieval found the answer; with 207 chunks of real statute the operative withdrawal provision sits at rank 14. A hollow measurement was replaced by a real one. Recorded in eval_results.md section 4c. Test count 117 to 122.
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
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.
Bumps langchain-community from 0.0.10 to 0.3.27.
Commits
bdf1cd3fix(langchain): update deps77c9819fix(text-splitters): update langchain-core version to 0.3.727f015b6fix(text-splitters): update lock for release71ad451Merge branch 'master' of github.com:langchain-ai/langchain2c42893fix(langchain): update langchain-core version to 0.3.720e139fbrelease(langchain): 0.3.27 (#32227)622bb05fix(langchain): class HTMLSemanticPreservingSplitter ignores the text inside ...56dde3afeat(langchain): v1 scaffolding (#32166)bd3d649release(core): 0.3.72 (#32214)fb5da83fix(core): Dereference Refs for pydantic schema fails in tool schema generati...You can trigger a rebase of this PR by commenting
@dependabot rebase.Dependabot commands and options
You can trigger Dependabot actions by commenting on this PR:
@dependabot rebasewill rebase this PR@dependabot recreatewill recreate this PR, overwriting any edits that have been made to it@dependabot show <dependency name> ignore conditionswill show all of the ignore conditions of the specified dependency@dependabot ignore this major versionwill close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)@dependabot ignore this minor versionwill close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)@dependabot ignore this dependencywill close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)You can disable automated security fix PRs for this repo from the Security Alerts page.