Skip to content

0.9.1 — SQL provenance fixes and reproducible tool comparisons

Choose a tag to compare

@sfy257000 sfy257000 released this 13 Nov 06:41
1535f89

This patch release corrects SQL access attribution and adds reproducible comparisons explaining when Local Codebase Indexer helps with SQL Server / C# maintenance.

Correctness fixes

  • Limit sp_executesql analysis to its SQL argument and statement, including quoted procedure names. Unrelated literals, parameter values, and similarly named columns no longer create false accesses.
  • Scope CTE names and UPDATE/DELETE aliases to their consuming statement, including omitted semicolons and UNION / INSERT SELECT patterns.
  • Recognize GO only as a batch separator line, preserving comma-joined sources after predicates such as a.go = b.id.
  • Keep SQL object definitions and suggestions restricted to SQL source files; application functions retain normal symbol navigation.
  • Respect C# / VB6 parameter and local shadowing, invalidate overwritten SQL bindings, and avoid treating query-parameter strings as executed SQL. Preserve earlier valid execution evidence.

Reproducible tool comparisons

English, Traditional Chinese, and Simplified Chinese READMEs now include measured results, trade-offs, and links to all raw query records. A new optional runner compares the unchanged retrieval questions with actual ripgrep candidate search plus standard BM25 ranking.

Retrieval suite rg + BM25 file Recall@5 Indexer Difference
MES-style synthetic, 42 questions 81.0% 91.7% +10.7 percentage points
Same questions with distractors 81.0% 90.5% +9.5 percentage points
Python standard library, 12 questions 100.0% 79.2% -20.8 percentage points

The indexer retains its default 5 result chunks; the baseline returns 5 files. Both are scored at file level. These are 54 unique project-authored questions and 96 executions including the repeated noisy suite, not independent production-accuracy measurements.

  • In the separate bundled synthetic navigation task, required evidence-file coverage increases from 3/5 to 5/5 (60% to 100%) versus one literal table-name search. Controller reachability remains probable, not runtime proof.
  • On noisy/stdlib retrieval suites, rendered output uses 94.9% / 96.9% fewer Unicode characters than full reads of the same selected files. Tiny synthetic files produce 2.3% more characters. These are not measured Claude-token savings or reductions against Grep snippets.
  • Richer navigation costs more time: recorded warm CLI medians are 244.5 ms versus 8.3 ms for a literal rg call, with 266.1 ms for the initial indexer call including indexing. Timings are machine-specific.
  • Serena and Claude Context features are documented with pinned upstream sources; no end-to-end Claude, Serena, or Claude Context superiority claim is made.

See the protocol, recorded summary, and all 96 query records. Recorded measurements retain their pre-release package-version label and source fingerprint.

Upgrade

python -m pip install --upgrade local-codebase-indexer==0.9.1

Index format increases from 9 to 10. The first index update automatically rebuilds older parsed data; existing projects do not need to delete their index manually.

Validation

187 tests pass, including 33 SQL/binding regressions and 7 benchmark-protocol tests added since 0.9.0. CI covers Linux and Windows on Python 3.10, 3.12, and 3.13, plus lint/type checks, installed-package validation, project evaluation, and comparison smoke runs. Retrieval questions, gold answers, and thresholds were not changed.

Changes: #4 and #5. Full changelog.