rag-your-code 1.5.1__tar.gz → 1.5.3__tar.gz

{rag_your_code-1.5.1/src/rag_your_code.egg-info → rag_your_code-1.5.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rag-your-code
-Version: 1.5.1
+Version: 1.5.3
 Summary: A local, explainable RAG index for codebases and coding agents
 Author: rag-your-code contributors
 License-Expression: MIT
@@ -224,7 +224,7 @@ A question with no lexical shortcut, asked of this repository:
 
 ````console
 $ rag-your-code search "where does it decide whether to answer at all" --limit 1
-[src/ragyourcode/search.py:117:Evidence] score=0.447
+[src/ragyourcode/search.py:117:Evidence] score=0.446
 The verdict on whether a question reached this index at all, kept separate from
 how results rank. ... 中文：判定一个提问究竟有没有够到索引的结论。...
 ```python
@@ -237,7 +237,7 @@ There is no string here to grep for: *decide* occurs nowhere in that
 declaration and matched nothing. What ranked it first is ordinary words —
 *answer*, *whether*, *where* — rare enough in this corpus to tell declarations
 apart. What the agent-written description adds is the other language:
-「在哪里判定一个提问有没有答案」 returns the same declaration first, at 0.395,
+「在哪里判定一个提问有没有答案」 returns the same declaration first, at 0.396,
 sharing not one character with its source.
 
 Now the case that motivated 1.0.0 — a question with no answer here at all:
@@ -261,12 +261,12 @@ it happens to use elsewhere.
                "matched_terms": ["job","leave","print"],
                "ubiquitous_terms": ["a","does","the","why"],
                "coverage": 0.5,    "min_coverage": 0.4,
-               "concentration": 0.1691, "min_concentration": 0.28,
+               "concentration": 0.1693, "min_concentration": 0.28,
                "applied_min_coverage": 0.4, "applied_min_concentration": 0.28,
                "hint": "..."}}
 ```
 
-Read `coverage: 0.5` against `concentration: 0.1691`. Half the distinctive
+Read `coverage: 0.5` against `concentration: 0.1693`. Half the distinctive
 words are here — `job`, `leave`, `print` — and spread thin enough that no
 declaration holds a fifth of what was asked, against a bar of 0.28. Before
 1.1.0 it came back with a confident-looking result.
@@ -366,18 +366,14 @@ a default can have.
 | refusing an unanswerable query | **0.016 ms** | 0.015 – 0.018 |
 | refusal cheaper than answering by | **~44×** | 40 – 47 |
 
-*Idle* is load-bearing: the same corpus at the same commit measured 0.99 ms
-median while a coverage run was in progress and 0.49 ms once it finished.
-This row read 0.49 ms in 1.5.0 at 601 units; a corpus 0.5% larger cannot
-explain 40%, and a repeat run here read 0.69 ms — the machine moved, not the code.
-
-Two significant figures and a spread, because that is the precision this has.
-Across five releases on the same idle machine the median has landed from 0.49
-to 1.44 ms and p95 from 0.85 to 7.34 ms — a band wider than any change the code
-has made to it. Releases before 1.3.0 published `0.83 ms / p95 1.68 ms` to
-three figures from an uncommitted script; both sit inside that band, which is
-the point — unfalsifiable rather than wrong. Refusal is cheap structurally: an
-unanswerable query touches only its distinctive words' posting lists, never ranking.
+*Idle* is load-bearing, and nothing in the report can see whether it was: the
+same corpus at the same commit measures about twice this median while another
+job is running. Two significant figures and a spread, because across releases
+that spread has been wider than any change the code has made to this number —
+which is why the row ships with its range and its corpus stamp rather than to
+three figures. Refusal is cheap structurally, not by tuning: an unanswerable
+query touches only its distinctive words' posting lists and never reaches
+ranking.
 
 **Scale**, synthetic 10,000-unit repository (500 files), re-measured in 1.5.0
 — the previous row of figures was optimistic by more than noise:
@@ -407,7 +403,7 @@ it again.
 Directional local measurements, not service levels — but each is a command
 rather than a memory, which two of them were not before. Each
 prints the corpus fingerprint beside its score; quote both or neither.
-[`benchmarks/README.md`](benchmarks/README.md) lists the seven scripts and what
+[`benchmarks/README.md`](benchmarks/README.md) lists the eight scripts and what
 each is for, and the corpus one of them grades is now carried here too.
 
 ## 7 · `rag-your-code search` vs a Grep loop
@@ -466,8 +462,8 @@ Qualifications, because the tables would otherwise flatter both sides:
   every file back in no order. It is also why Grep declines nine of the seventy
   questions here — no word was left that this corpus does not use everywhere.
 - **Payload is counted in characters on both sides.** Grep hands back roughly
-  19,300 characters per question it answers here, unranked and without spans,
-  against 10,300 ranked and capped by `search.max_chars` — a factor of 1.9,
+  19,500 characters per question it answers here, unranked and without spans,
+  against 10,200 ranked and capped by `search.max_chars` — a factor of 1.9,
   5.5 on Flask and 5.1 on cobra, where a framework repeats its vocabulary
   across files and Grep cannot rank what it finds. 1.4.1 changed what fits in
   that cap: the block had been reprinting the docstring the code below already
@@ -731,11 +727,11 @@ by which the author's docstring reaches the weight-3 description field — so
 writing one demotes it to the weight-1 body. On `parser.py::_generic_units` a
 long description cost one graded question and a short one cost three;
 appending the docstring to every description instead cost the 1.5.0 corpus
-0.443 → 0.414 hit@1. `describe.skip` records the decision.
+0.429 → 0.414 hit@1. `describe.skip` records the decision.
 
-**The vectors are 72.1% of the index and earn at most two questions** under the
-default embedder — 74.8% Flask, 79.7% cobra. Ablating costs A one, gains B one
-and C two at hit@3, moves E none; that storage is what an optional model needs.
+**The vectors are 72.1% of the index and earn at most two questions** on any
+ruler, in either direction, under the default embedder — 74.8% Flask, 79.7%
+cobra. Kept: that same storage is what an optional model needs.
 
 **`search.vector_recall` scans every vector per query** — under a semantic
 embedder. The default hash never widens at all. Affordable at the measured
@@ -767,7 +763,8 @@ that rots. `pytest --cov=ragyourcode` is the command behind the coverage one.
 CI runs Python 3.10–3.13 on Linux and Windows, installs the built wheel into a
 clean environment and runs the documented CLI end to end — `bootstrap` through
 `describe promote` — plus the skill's own install line verbatim, and grades
-every ruler including both vendored corpora.
+rulers A, C, D and E across this repository and both vendored corpora. Ruler B,
+the cold parse of this repository, is a local command.
 
 - [docs/FLOW.md](docs/FLOW.md) — the whole thing in four diagrams
 - [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) — how each stage works and why

{rag_your_code-1.5.1 → rag_your_code-1.5.3}/README.md

@@ -194,7 +194,7 @@ A question with no lexical shortcut, asked of this repository:
 
 ````console
 $ rag-your-code search "where does it decide whether to answer at all" --limit 1
-[src/ragyourcode/search.py:117:Evidence] score=0.447
+[src/ragyourcode/search.py:117:Evidence] score=0.446
 The verdict on whether a question reached this index at all, kept separate from
 how results rank. ... 中文：判定一个提问究竟有没有够到索引的结论。...
 ```python
@@ -207,7 +207,7 @@ There is no string here to grep for: *decide* occurs nowhere in that
 declaration and matched nothing. What ranked it first is ordinary words —
 *answer*, *whether*, *where* — rare enough in this corpus to tell declarations
 apart. What the agent-written description adds is the other language:
-「在哪里判定一个提问有没有答案」 returns the same declaration first, at 0.395,
+「在哪里判定一个提问有没有答案」 returns the same declaration first, at 0.396,
 sharing not one character with its source.
 
 Now the case that motivated 1.0.0 — a question with no answer here at all:
@@ -231,12 +231,12 @@ it happens to use elsewhere.
                "matched_terms": ["job","leave","print"],
                "ubiquitous_terms": ["a","does","the","why"],
                "coverage": 0.5,    "min_coverage": 0.4,
-               "concentration": 0.1691, "min_concentration": 0.28,
+               "concentration": 0.1693, "min_concentration": 0.28,
                "applied_min_coverage": 0.4, "applied_min_concentration": 0.28,
                "hint": "..."}}
 ```
 
-Read `coverage: 0.5` against `concentration: 0.1691`. Half the distinctive
+Read `coverage: 0.5` against `concentration: 0.1693`. Half the distinctive
 words are here — `job`, `leave`, `print` — and spread thin enough that no
 declaration holds a fifth of what was asked, against a bar of 0.28. Before
 1.1.0 it came back with a confident-looking result.
@@ -336,18 +336,14 @@ a default can have.
 | refusing an unanswerable query | **0.016 ms** | 0.015 – 0.018 |
 | refusal cheaper than answering by | **~44×** | 40 – 47 |
 
-*Idle* is load-bearing: the same corpus at the same commit measured 0.99 ms
-median while a coverage run was in progress and 0.49 ms once it finished.
-This row read 0.49 ms in 1.5.0 at 601 units; a corpus 0.5% larger cannot
-explain 40%, and a repeat run here read 0.69 ms — the machine moved, not the code.
-
-Two significant figures and a spread, because that is the precision this has.
-Across five releases on the same idle machine the median has landed from 0.49
-to 1.44 ms and p95 from 0.85 to 7.34 ms — a band wider than any change the code
-has made to it. Releases before 1.3.0 published `0.83 ms / p95 1.68 ms` to
-three figures from an uncommitted script; both sit inside that band, which is
-the point — unfalsifiable rather than wrong. Refusal is cheap structurally: an
-unanswerable query touches only its distinctive words' posting lists, never ranking.
+*Idle* is load-bearing, and nothing in the report can see whether it was: the
+same corpus at the same commit measures about twice this median while another
+job is running. Two significant figures and a spread, because across releases
+that spread has been wider than any change the code has made to this number —
+which is why the row ships with its range and its corpus stamp rather than to
+three figures. Refusal is cheap structurally, not by tuning: an unanswerable
+query touches only its distinctive words' posting lists and never reaches
+ranking.
 
 **Scale**, synthetic 10,000-unit repository (500 files), re-measured in 1.5.0
 — the previous row of figures was optimistic by more than noise:
@@ -377,7 +373,7 @@ it again.
 Directional local measurements, not service levels — but each is a command
 rather than a memory, which two of them were not before. Each
 prints the corpus fingerprint beside its score; quote both or neither.
-[`benchmarks/README.md`](benchmarks/README.md) lists the seven scripts and what
+[`benchmarks/README.md`](benchmarks/README.md) lists the eight scripts and what
 each is for, and the corpus one of them grades is now carried here too.
 
 ## 7 · `rag-your-code search` vs a Grep loop
@@ -436,8 +432,8 @@ Qualifications, because the tables would otherwise flatter both sides:
   every file back in no order. It is also why Grep declines nine of the seventy
   questions here — no word was left that this corpus does not use everywhere.
 - **Payload is counted in characters on both sides.** Grep hands back roughly
-  19,300 characters per question it answers here, unranked and without spans,
-  against 10,300 ranked and capped by `search.max_chars` — a factor of 1.9,
+  19,500 characters per question it answers here, unranked and without spans,
+  against 10,200 ranked and capped by `search.max_chars` — a factor of 1.9,
   5.5 on Flask and 5.1 on cobra, where a framework repeats its vocabulary
   across files and Grep cannot rank what it finds. 1.4.1 changed what fits in
   that cap: the block had been reprinting the docstring the code below already
@@ -701,11 +697,11 @@ by which the author's docstring reaches the weight-3 description field — so
 writing one demotes it to the weight-1 body. On `parser.py::_generic_units` a
 long description cost one graded question and a short one cost three;
 appending the docstring to every description instead cost the 1.5.0 corpus
-0.443 → 0.414 hit@1. `describe.skip` records the decision.
+0.429 → 0.414 hit@1. `describe.skip` records the decision.
 
-**The vectors are 72.1% of the index and earn at most two questions** under the
-default embedder — 74.8% Flask, 79.7% cobra. Ablating costs A one, gains B one
-and C two at hit@3, moves E none; that storage is what an optional model needs.
+**The vectors are 72.1% of the index and earn at most two questions** on any
+ruler, in either direction, under the default embedder — 74.8% Flask, 79.7%
+cobra. Kept: that same storage is what an optional model needs.
 
 **`search.vector_recall` scans every vector per query** — under a semantic
 embedder. The default hash never widens at all. Affordable at the measured
@@ -737,7 +733,8 @@ that rots. `pytest --cov=ragyourcode` is the command behind the coverage one.
 CI runs Python 3.10–3.13 on Linux and Windows, installs the built wheel into a
 clean environment and runs the documented CLI end to end — `bootstrap` through
 `describe promote` — plus the skill's own install line verbatim, and grades
-every ruler including both vendored corpora.
+rulers A, C, D and E across this repository and both vendored corpora. Ruler B,
+the cold parse of this repository, is a local command.
 
 - [docs/FLOW.md](docs/FLOW.md) — the whole thing in four diagrams
 - [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) — how each stage works and why

{rag_your_code-1.5.1 → rag_your_code-1.5.3}/pyproject.toml

@@ -6,7 +6,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "rag-your-code"
-version = "1.5.1"
+version = "1.5.3"
 description = "A local, explainable RAG index for codebases and coding agents"
 readme = "README.md"
 requires-python = ">=3.10"

{rag_your_code-1.5.1 → rag_your_code-1.5.3/src/rag_your_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rag-your-code
-Version: 1.5.1
+Version: 1.5.3
 Summary: A local, explainable RAG index for codebases and coding agents
 Author: rag-your-code contributors
 License-Expression: MIT
@@ -224,7 +224,7 @@ A question with no lexical shortcut, asked of this repository:
 
 ````console
 $ rag-your-code search "where does it decide whether to answer at all" --limit 1
-[src/ragyourcode/search.py:117:Evidence] score=0.447
+[src/ragyourcode/search.py:117:Evidence] score=0.446
 The verdict on whether a question reached this index at all, kept separate from
 how results rank. ... 中文：判定一个提问究竟有没有够到索引的结论。...
 ```python
@@ -237,7 +237,7 @@ There is no string here to grep for: *decide* occurs nowhere in that
 declaration and matched nothing. What ranked it first is ordinary words —
 *answer*, *whether*, *where* — rare enough in this corpus to tell declarations
 apart. What the agent-written description adds is the other language:
-「在哪里判定一个提问有没有答案」 returns the same declaration first, at 0.395,
+「在哪里判定一个提问有没有答案」 returns the same declaration first, at 0.396,
 sharing not one character with its source.
 
 Now the case that motivated 1.0.0 — a question with no answer here at all:
@@ -261,12 +261,12 @@ it happens to use elsewhere.
                "matched_terms": ["job","leave","print"],
                "ubiquitous_terms": ["a","does","the","why"],
                "coverage": 0.5,    "min_coverage": 0.4,
-               "concentration": 0.1691, "min_concentration": 0.28,
+               "concentration": 0.1693, "min_concentration": 0.28,
                "applied_min_coverage": 0.4, "applied_min_concentration": 0.28,
                "hint": "..."}}
 ```
 
-Read `coverage: 0.5` against `concentration: 0.1691`. Half the distinctive
+Read `coverage: 0.5` against `concentration: 0.1693`. Half the distinctive
 words are here — `job`, `leave`, `print` — and spread thin enough that no
 declaration holds a fifth of what was asked, against a bar of 0.28. Before
 1.1.0 it came back with a confident-looking result.
@@ -366,18 +366,14 @@ a default can have.
 | refusing an unanswerable query | **0.016 ms** | 0.015 – 0.018 |
 | refusal cheaper than answering by | **~44×** | 40 – 47 |
 
-*Idle* is load-bearing: the same corpus at the same commit measured 0.99 ms
-median while a coverage run was in progress and 0.49 ms once it finished.
-This row read 0.49 ms in 1.5.0 at 601 units; a corpus 0.5% larger cannot
-explain 40%, and a repeat run here read 0.69 ms — the machine moved, not the code.
-
-Two significant figures and a spread, because that is the precision this has.
-Across five releases on the same idle machine the median has landed from 0.49
-to 1.44 ms and p95 from 0.85 to 7.34 ms — a band wider than any change the code
-has made to it. Releases before 1.3.0 published `0.83 ms / p95 1.68 ms` to
-three figures from an uncommitted script; both sit inside that band, which is
-the point — unfalsifiable rather than wrong. Refusal is cheap structurally: an
-unanswerable query touches only its distinctive words' posting lists, never ranking.
+*Idle* is load-bearing, and nothing in the report can see whether it was: the
+same corpus at the same commit measures about twice this median while another
+job is running. Two significant figures and a spread, because across releases
+that spread has been wider than any change the code has made to this number —
+which is why the row ships with its range and its corpus stamp rather than to
+three figures. Refusal is cheap structurally, not by tuning: an unanswerable
+query touches only its distinctive words' posting lists and never reaches
+ranking.
 
 **Scale**, synthetic 10,000-unit repository (500 files), re-measured in 1.5.0
 — the previous row of figures was optimistic by more than noise:
@@ -407,7 +403,7 @@ it again.
 Directional local measurements, not service levels — but each is a command
 rather than a memory, which two of them were not before. Each
 prints the corpus fingerprint beside its score; quote both or neither.
-[`benchmarks/README.md`](benchmarks/README.md) lists the seven scripts and what
+[`benchmarks/README.md`](benchmarks/README.md) lists the eight scripts and what
 each is for, and the corpus one of them grades is now carried here too.
 
 ## 7 · `rag-your-code search` vs a Grep loop
@@ -466,8 +462,8 @@ Qualifications, because the tables would otherwise flatter both sides:
   every file back in no order. It is also why Grep declines nine of the seventy
   questions here — no word was left that this corpus does not use everywhere.
 - **Payload is counted in characters on both sides.** Grep hands back roughly
-  19,300 characters per question it answers here, unranked and without spans,
-  against 10,300 ranked and capped by `search.max_chars` — a factor of 1.9,
+  19,500 characters per question it answers here, unranked and without spans,
+  against 10,200 ranked and capped by `search.max_chars` — a factor of 1.9,
   5.5 on Flask and 5.1 on cobra, where a framework repeats its vocabulary
   across files and Grep cannot rank what it finds. 1.4.1 changed what fits in
   that cap: the block had been reprinting the docstring the code below already
@@ -731,11 +727,11 @@ by which the author's docstring reaches the weight-3 description field — so
 writing one demotes it to the weight-1 body. On `parser.py::_generic_units` a
 long description cost one graded question and a short one cost three;
 appending the docstring to every description instead cost the 1.5.0 corpus
-0.443 → 0.414 hit@1. `describe.skip` records the decision.
+0.429 → 0.414 hit@1. `describe.skip` records the decision.
 
-**The vectors are 72.1% of the index and earn at most two questions** under the
-default embedder — 74.8% Flask, 79.7% cobra. Ablating costs A one, gains B one
-and C two at hit@3, moves E none; that storage is what an optional model needs.
+**The vectors are 72.1% of the index and earn at most two questions** on any
+ruler, in either direction, under the default embedder — 74.8% Flask, 79.7%
+cobra. Kept: that same storage is what an optional model needs.
 
 **`search.vector_recall` scans every vector per query** — under a semantic
 embedder. The default hash never widens at all. Affordable at the measured
@@ -767,7 +763,8 @@ that rots. `pytest --cov=ragyourcode` is the command behind the coverage one.
 CI runs Python 3.10–3.13 on Linux and Windows, installs the built wheel into a
 clean environment and runs the documented CLI end to end — `bootstrap` through
 `describe promote` — plus the skill's own install line verbatim, and grades
-every ruler including both vendored corpora.
+rulers A, C, D and E across this repository and both vendored corpora. Ruler B,
+the cold parse of this repository, is a local command.
 
 - [docs/FLOW.md](docs/FLOW.md) — the whole thing in four diagrams
 - [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) — how each stage works and why

{rag_your_code-1.5.1 → rag_your_code-1.5.3}/src/ragyourcode/__init__.py

@@ -3,4 +3,4 @@
 from .models import CodeUnit, SearchResult
 
 __all__ = ["CodeUnit", "SearchResult"]
-__version__ = "1.5.1"
+__version__ = "1.5.3"

{rag_your_code-1.5.1 → rag_your_code-1.5.3}/tests/test_metadata.py

@@ -345,7 +345,7 @@ def test_every_benchmark_script_is_listed_in_its_own_index():
     Both directions matter. A script absent from the index is one nobody knows
     to run; a command in the index naming a script that does not exist is the
     install-line defect this repository shipped twice. Discovery by glob is
-    what keeps the seventh script from being the one nothing checks.
+    what keeps the newest script from being the one nothing checks.
     """
     index = (ROOT / "benchmarks" / "README.md").read_text(encoding="utf-8")
     assert BENCHMARKS, "benchmarks/ holds no scripts; this guard would pass vacuously"