feat: v1.6.0 search quality improvements #26
Conversation
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 08d243d956
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
PatrickSys
force-pushed
the
feat/v1.6.0-search-quality
branch
from
08d243d to
91b6633
Compare
PatrickSys
force-pushed
the
feat/v1.6.0-search-quality
branch
from
91b6633 to
bc836d1
Compare
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: bc836d15be
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
PatrickSys
force-pushed
the
feat/v1.6.0-search-quality
branch
from
bc836d1 to
26aa3f7
Compare
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 26aa3f7f2a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
PatrickSys
force-pushed
the
feat/v1.6.0-search-quality
branch
from
26aa3f7 to
f89e7a3
Compare
…nker - Reciprocal Rank Fusion (k=60, rank-based scoring) - Intent-aware query classification (5 types) - Hard test-file filtering (0% spec contamination) - Import-graph proximity reranking - File-level deduplication - Chunk optimization (100→50 lines, 0 overlap) - Embedding upgrade (@huggingface/transformers v3) BREAKING CHANGE: Re-indexing required due to model and chunking changes
PatrickSys
force-pushed
the
feat/v1.6.0-search-quality
branch
from
f89e7a3 to
5346766
Compare
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 534676663a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const reranked = scored.map(({ result, crossScore }) => ({ | ||
| ...result, | ||
| // Normalize cross-encoder score to 0-1 range for the top slot | ||
| score: crossScore |
There was a problem hiding this comment.
This writes raw cross-encoder logits into SearchResult.score, but downstream quality logic (assessSearchQuality) interprets scores as normalized confidence values (for example, low-score and separation thresholds). When reranking is triggered on ambiguous queries (enabled by default), negative or unbounded logits can incorrectly downgrade otherwise good results to low_confidence and distort confidence signals returned to users. Convert reranker outputs to a bounded confidence scale (or store them separately) before replacing score.
Useful? React with 👍 / 👎.
| if (this.isTemplateOrStyleFile(chunk.filePath)) { | ||
| combinedScore *= 0.75; |
There was a problem hiding this comment.
The template/style penalty currently applies to every query, even though the comment says it is for behavioral queries. This means style- or template-focused searches (for example, layout/theme/CSS questions) always get a 25% penalty on the most relevant .html/.scss files, which can push unrelated implementation files above the correct targets and reduce retrieval quality.
Useful? React with 👍 / 👎.
1 participant
Summary
@xenova/transformersv2 →@huggingface/transformersv3 (granite model, 8192 context)BREAKING CHANGE
Re-indexing required after upgrade. Existing
.codebase-context/indices from v1.5.x are incompatible.Test plan
pnpm buildsucceedspnpm test— 135/135 passing