pyvectorhound 1.3.0__tar.gz → 1.5.0__tar.gz

{pyvectorhound-1.3.0 → pyvectorhound-1.5.0}/.github/workflows/ci.yml

@@ -13,7 +13,7 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - uses: dtolnay/rust-toolchain@v1
+      - uses: dtolnay/rust-toolchain@stable
 
       - uses: Swatinem/rust-cache@v2
 
@@ -32,7 +32,7 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - uses: dtolnay/rust-toolchain@v1
+      - uses: dtolnay/rust-toolchain@stable
 
       - uses: Swatinem/rust-cache@v2
 
@@ -40,13 +40,7 @@ jobs:
         with:
           python-version: ${{ matrix.python-version }}
 
-      - name: Install maturin
-        run: pip install "maturin>=1.7,<2.0"
-
-      - name: Build and install extension module
-        run: maturin develop --release
-
-      - name: Install dependencies
+      - name: Install package + dev dependencies
         run: pip install -e ".[dev]"
 
       - name: Run pytest

pyvectorhound-1.5.0/CHANGELOG.md

@@ -0,0 +1,192 @@
+# Changelog
+
+All notable changes to PyVectorHound are documented in this file.
+
+## [1.5.0]
+
+### Added
+
+- **int8 scalar quantization** (`src/quantization.rs`) — per-vector min/max
+  linear quantization to `i8`, with `quantize`/`dequantize` (and batch
+  variants) exposed via PyO3. 4x memory reduction vs `f32` for large local
+  vector indices, with measured reconstruction error (0.0039 max on a
+  384-dim `[-1,1]` test vector) within the theoretical 8-bit bound
+  (`range/255`). SIMD/GPU acceleration is a separate, larger effort and
+  intentionally out of scope here.
+
+## [1.4.0]
+
+### Added
+
+- **LLM-as-judge faithfulness / semantic contradiction scoring.** `Diagnosis`
+  accepts optional `document_texts` (doc_id -> text) and `llm_judge_fn`
+  (`(query, doc_texts) -> dict`) — mirrors the existing `embed_fn` pattern,
+  no LLM client bundled. Adds a `faithfulness` component to
+  `metrics()`/`hunt()`/`recommendations()`/`root_cause()` that catches
+  retrieved documents that are embedding-similar but semantically
+  contradict or are irrelevant to the query, a failure mode distance
+  metrics alone can't see. Reports `UNKNOWN` honestly when the required
+  inputs aren't supplied.
+- **`Hound.diagnose_batch()`** — runs many queries concurrently on a thread
+  pool instead of serially, for large-scale evaluation runs. `embed_fn`
+  and `llm_judge_fn` calls now retry with exponential backoff + jitter
+  (`pyvectorhound/_retry.py`).
+- **Dynamic threshold calibration.** `scorer.py` and `diagnosis.py`'s
+  GOOD/MODERATE/WEAK quality-status cutoffs now classify relative to a
+  tracked `TrendAnalyzer` baseline (z-score based) when one is available,
+  instead of a fixed numeric cutoff — so status doesn't silently drift
+  when you swap embedding models. Falls back to the original fixed
+  cutoffs on cold start (no baseline yet). `Hound.diagnose()` and
+  `Hound.quality_scorer()` wire this up automatically.
+
+## [1.3.3] - Unreleased
+
+This release reconciles two independent fix passes that diverged from the
+same base commit (`165cc80`) without knowledge of each other:
+
+1. The repo owner's own prior work (now `1.3.2`, below): the v1.3 retrieval
+   ranking feature, a real Milvus `get_embeddings`, and a fix for the
+   `_core` extension / fabricated-diagnosis bug.
+2. A separate remediation pass, done from a stale local clone that had not
+   pulled (1), which fixed the *same* `_core`/fabrication bug independently
+   (briefly published to PyPI as `1.3.0` before this reconciliation) and
+   also fixed a number of other, non-overlapping fabrication bugs the
+   `1.3.2` pass didn't touch.
+
+Where both passes fixed the same thing (`diagnosis.py`, `scorer.py`,
+`src/lib.rs`'s `_core` symbol, the `pyhound` import), `1.3.2`'s
+implementation was kept as authoritative and the duplicate was discarded.
+Everything below is what pass (2) fixed that pass (1) did not; it's applied
+on top of `1.3.2` unchanged.
+
+### Fixed
+- `Hound.diagnose()` fabricated the query embedding: a query with no
+  `query_embedding` got `np.random.randn(768)`, silently producing a
+  meaningless diagnosis dressed up with real-looking metrics. It now
+  requires a `query_embedding` or a configured `embed_fn` and raises a
+  clear `ValueError` otherwise. **Breaking change** for callers that relied
+  on the old silent random-vector fallback; `tests/test_hound.py` was
+  updated to assert the new behavior (and a new `embed_fn`-path test) in
+  place of the old test that asserted the fabricated vector was generated.
+- `Hound.__init__` eagerly called `adapter.connect()`, requiring the
+  optional client library and a live server just to construct a `Hound`.
+  Connection is now lazy, matching every adapter method's own
+  lazy-connect behavior.
+- Database adapters' `search()` (Qdrant, Chroma, Milvus, pgvector,
+  Weaviate) returned the *query* embedding as every result's `"embedding"`
+  field instead of the real per-document vector, corrupting any downstream
+  embedding-space analysis that reads it directly from search results.
+  They now fetch and return the actual per-document embedding where the
+  underlying client supports it. (Milvus's `search()` additionally now
+  requests `vector_field`/`id_field` per `1.3.2`'s configurable field
+  names, alongside this fix.)
+- `pydantic` is used in `database.py` (`SecretStr`) but was missing from
+  `pyproject.toml` dependencies, so a clean install could fail on import
+  if pydantic wasn't already present transitively. Added as a dependency.
+  Also moved the `from pydantic import SecretStr` statement below the
+  module docstring -- it previously preceded it, which silently discarded
+  the module's `__doc__`.
+- Base and optional dependencies were pinned to exact old versions (e.g.
+  `numpy==1.20.0`, `black==23.0.0`) with no compatible wheels for some of
+  the Python versions this package claims to support (3.8-3.12), breaking
+  `pip install .[dev]`. Relaxed to compatible ranges.
+- CI's Python test job referenced a nonexistent `python/` subdirectory and
+  never built the Rust extension via `maturin`, so it always ran against
+  whatever `_core` state happened to exist (or silently skipped tests
+  entirely). It now runs `maturin develop --release` before installing
+  dependencies and running pytest, at the actual repo root.
+- Adapter `ImportError` messages told users to `pip install pyhound[...]`
+  -- `pyhound` is this project's old name and is not the installable
+  package; corrected to `pyvectorhound[...]` across all five adapters.
+- `ModelComparison`/`Hound.compare_models()` fabricated quality scores
+  (`f1_score`, `ndcg`, etc.) derived from `hash(model_name)`. It now
+  reports only real, published cost/latency metadata unless a `quality_fn`
+  is supplied for real, measured quality numbers; unmeasured quality is
+  reported as `None` with `quality_measured: False` rather than a fake
+  number.
+- `Hound.compare_metrics()` and `Hound.detect_drift()` returned fixed
+  placeholder output (zeros / `"No action needed"`) regardless of input,
+  with no backing data store. They now raise `NotImplementedError`
+  pointing at the real, tested alternative (`track_metric()` +
+  `get_trend_report()` / `TrendAnalyzer.detect_drift()`).
+- Removed `tests/test_vector_operations.py`: it tested `VectorStore` and
+  `EmbeddingManager` classes that do not exist anywhere in this codebase
+  and are outside this package's scope (diagnosing failures in databases
+  you already run, not being one).
+- Removed stale `README.md.bak` and `pyvectorhound/__init__.py.bak` files
+  left over from a previous edit.
+- `pyvectorhound/diagnosis.py`'s `hunt()` report header still read
+  "PyHound Diagnosis Report" after the project's rename; corrected to
+  "PyVectorHound", with the matching test assertion updated.
+- The CLI (`cli.py`) and REST server (`server.py`) never threaded
+  `query_embedding`/`embed_fn` through to `Hound.diagnose()`, so both would
+  hit the `ValueError` above unconditionally once the random-embedding
+  fallback was removed. Both now accept and forward a precomputed
+  `query_embedding`.
+- `README.md` and `examples/retrieval_debug.py` described a fictional API
+  (`VectorHound`, `hound.validate_quality()`, `quality_score()`, DB
+  support for Pinecone/Elasticsearch that isn't implemented) instead of
+  the real `Hound`/`Diagnosis` classes. Rewritten to match the actual,
+  reconciled API, including `embed_fn` and lazy connection.
+
+## [1.3.2] - Unreleased
+
+### Fixed
+- The compiled Rust scoring engine (`pyvectorhound._core`) could never actually load, for two independent reasons: (1) the `#[pymodule]` function was named `pyhound_core` in Rust while `Cargo.toml`/`pyproject.toml` both expect the extension's init symbol to be `_core`; (2) `diagnosis.py` and `scorer.py` imported it as `from pyhound import _core` -- `pyhound` was this project's old name before it was renamed to `pyvectorhound`, and that import was never updated. Both are fixed; `_core` now loads and its isotropy/coverage/distinctiveness/quality-score functions are actually reachable for the first time.
+- `Diagnosis._analyze_embedding()` returned hardcoded placeholder numbers (isotropy 0.72, coverage 0.85, distinctiveness 0.68) and an unconditional `"GOOD"` status regardless of the actual retrieved documents. It now fetches the real per-document embeddings via `adapter.get_embeddings()` and computes real metrics via `_core`, honestly reporting `UNKNOWN` when there isn't enough real data to compute them (fewer than 2 embeddings, no adapter, or `_core` unavailable) rather than a fabricated score.
+- `Diagnosis._analyze_vector_search()`'s MRR was a hardcoded `0.85 if any relevant doc was retrieved, else 0.0`, ignoring rank entirely. It's now a real reciprocal-rank calculation based on the position of the first relevant result.
+- `Diagnosis._analyze_bm25()` and `_analyze_reranker()` unconditionally reported `"GOOD"` with fabricated precision/recall/calibration/NDCG numbers even though no BM25 index or reranker-score subsystem exists anywhere in this codebase. They now honestly report `UNKNOWN` with an explanation, instead of a fake positive verdict. `root_cause()`/`recommendations()` updated so they no longer misinterpret an honest `UNKNOWN` as a real quality problem.
+- `QualityScorer.score()` computed a real `overall` score via `_core` but returned hardcoded `isotropy`/`coverage`/`distinctiveness`/`query_relevance` values alongside it regardless of the input embedding. It now samples real nearest-neighbor embeddings from the corpus via the adapter and computes all four metrics for real (or reports `UNKNOWN` if no adapter/corpus data is available, rather than a fake middling score).
+- `QualityScorer.corpus_health()` returned fixed `avg_isotropy`/`avg_coverage`/`avg_distinctiveness`/`drift`/`trend` values regardless of the actual corpus. It now samples real corpus embeddings and computes real quality metrics; `drift`/`trend` honestly report as not-measured (this class doesn't persist historical snapshots) instead of a fabricated `"stable"`.
+- `QualityScorer.trend_analysis()` returned a fixed `"stable"` direction and `0.02` magnitude for any date range, without looking at any data. It now honestly reports that no historical tracking is wired up, and points callers at `TrendAnalyzer` (which does real time-series tracking) instead.
+- `__version__` in `pyvectorhound/__init__.py` was hardcoded to `"1.2.1"` and had drifted from the actual package version; now reads it from installed package metadata.
+
+### Added
+- Real test coverage in `tests/test_diagnosis.py` and new `tests/test_scorer.py` asserting the above methods respond to real input data instead of returning fixed constants.
+
+### Fixed
+- `MilvusAdapter.get_embeddings()` was a stub that unconditionally returned an empty dict, discarding the requested document IDs instead of querying Milvus. It now retrieves the real stored vectors via `MilvusClient.get()`, with configurable `vector_field`/`id_field` names to match the caller's collection schema.
+- `MilvusAdapter.search()` used a hardcoded `"id"` output field instead of the configurable `id_field`.
+- `pyproject.toml` was still at 1.2.1 while `Cargo.toml` had already moved to 1.3.0 in a prior commit; synced both.
+
+### Added
+- Real test coverage for `pyvectorhound/database.py`'s adapters (`MilvusAdapter`, `QdrantAdapter`, `ChromaAdapter`), which previously had zero tests — only the separate mock-based `db_adapters.py` interface was tested.
+- Replaced two stub tests in `tests/test_hound.py` (`test_unsupported_db`, `test_diagnose`) that were empty `pass` statements with `# TODO` comments — they now actually exercise `Hound`'s real error path and its real adapter-search-to-Diagnosis flow.
+
+## [0.1.1] - 2025-01-20
+
+### Added
+- Component-level diagnostics for retrieval pipelines
+- Root cause analysis for embedding failures
+- Recommendations engine with ROI estimates
+- Support for Elasticsearch, Weaviate, and Pinecone
+- BM25 keyword search diagnostics
+- Reranker performance analysis
+
+### Fixed
+- Memory efficiency improvements in large batch processing
+- Vector similarity computation accuracy
+
+## [0.1.0] - 2024-12-15
+
+### Added
+- Initial release
+- Core retrieval diagnostics engine
+- Embedding quality assessment
+- Vector search performance profiling
+- MIT license
+
+### Known Issues
+- Streaming RAG pipelines not yet supported
+- LLamaIndex integration in progress
+
+## [Unreleased]
+
+### Planned
+- Streaming RAG support
+- LLamaIndex integration
+- Custom metric definitions
+- Dashboard UI
+- API server mode
+
+For details, see [ROADMAP.md](../ROADMAP.md).

{pyvectorhound-1.3.0 → pyvectorhound-1.5.0}/Cargo.lock

@@ -37,9 +37,9 @@ checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
 
 [[package]]
 name = "libc"
-version = "0.2.189"
+version = "0.2.186"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3eaf3ede3fee6db1a4c2ee091bf8a8b4dccdc6d17f656fb07896ee72867612f2"
+checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66"
 
 [[package]]
 name = "matrixmultiply"
@@ -146,9 +146,9 @@ dependencies = [
 
 [[package]]
 name = "proc-macro2"
-version = "1.0.107"
+version = "1.0.106"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "985e7ec9bb745e6ce6535b544d84d6cd6f7ad8bd711c398938ae983b91a766d9"
+checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934"
 dependencies = [
  "unicode-ident",
 ]
@@ -200,7 +200,7 @@ dependencies = [
  "proc-macro2",
  "pyo3-macros-backend",
  "quote",
- "syn 2.0.119",
+ "syn",
 ]
 
 [[package]]
@@ -213,12 +213,12 @@ dependencies = [
  "proc-macro2",
  "pyo3-build-config",
  "quote",
- "syn 2.0.119",
+ "syn",
 ]
 
 [[package]]
 name = "pyvectorhound"
-version = "1.3.0"
+version = "1.5.0"
 dependencies = [
  "numpy",
  "pyo3",
@@ -228,9 +228,9 @@ dependencies = [
 
 [[package]]
 name = "quote"
-version = "1.0.47"
+version = "1.0.46"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1fbf4db142a473a8d80c26bbf18454ed458bf8d26c8219c331daecfdbd079001"
+checksum = "dfbc457d0c7a0759a614551b11a6409e5951f6c7537be1f1b7682b9ae9230368"
 dependencies = [
  "proc-macro2",
 ]
@@ -255,9 +255,9 @@ checksum = "cf54715a573b99ac80df0bc206da022bcd442c974952c7b9720069370852e21f"
 
 [[package]]
 name = "serde"
-version = "1.0.229"
+version = "1.0.228"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4148590afebada386688f18773da617792bf2ef03ffc1e4cbd2b1d45b023e0ba"
+checksum = "9a8e94ea7f378bd32cbbd37198a4a91436180c5bb472411e48b5ec2e2124ae9e"
 dependencies = [
  "serde_core",
  "serde_derive",
@@ -265,29 +265,29 @@ dependencies = [
 
 [[package]]
 name = "serde_core"
-version = "1.0.229"
+version = "1.0.228"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "67dca2c9c51e58a4791a4b1ed58308b39c64224d349a935ab5039aa360942a48"
+checksum = "41d385c7d4ca58e59fc732af25c3983b67ac852c1a25000afe1175de458b67ad"
 dependencies = [
  "serde_derive",
 ]
 
 [[package]]
 name = "serde_derive"
-version = "1.0.229"
+version = "1.0.228"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e7a5d71263a5a7d47b41f6b3f06ba276f10cc18b0931f1799f710578e2309348"
+checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 3.0.3",
+ "syn",
 ]
 
 [[package]]
 name = "serde_json"
-version = "1.0.151"
+version = "1.0.150"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c841b55ecdae098c80dcae9cf767f6f8a0c2cdb3416bbef72181df4d0fe73f14"
+checksum = "e8014e44b4736ed0538adeecded0fce2a272f22dc9578a7eb6b2d9993c74cfb9"
 dependencies = [
  "itoa",
  "memchr",
@@ -307,17 +307,6 @@ dependencies = [
  "unicode-ident",
 ]
 
-[[package]]
-name = "syn"
-version = "3.0.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "53e9bae58849f64dfa4f5d5ae372c8341f7305f82a3868709269343628b659a3"
-dependencies = [
- "proc-macro2",
- "quote",
- "unicode-ident",
-]
-
 [[package]]
 name = "target-lexicon"
 version = "0.12.16"

{pyvectorhound-1.3.0 → pyvectorhound-1.5.0}/Cargo.toml

@@ -1,11 +1,10 @@
 [package]
 name = "pyvectorhound"
-version = "1.3.0"
+version = "1.5.0"
 edition = "2021"
 authors = ["Georgi Mammen Mullassery <mullassery@gmail.com>"]
 license = "Proprietary"
 description = "Vector search diagnostics engine for RAG/LLM systems"
-readme = "README.md"
 
 [dependencies]
 pyo3 = { version = "0.23", features = ["extension-module"] }