vector-engine 1.0.0__py3-none-any.whl

vector_engine-1.0.0.dist-info/METADATA

@@ -0,0 +1,342 @@
+Metadata-Version: 2.4
+Name: vector-engine
+Version: 1.0.0
+Summary: ML-first vector computation and retrieval engine.
+Author: Vector Engine Contributors
+License-Expression: MIT
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.24
+Provides-Extra: faiss
+Requires-Dist: faiss-cpu>=1.7.4; (platform_system != "Darwin" or platform_machine != "arm64") and extra == "faiss"
+Provides-Extra: ml
+Requires-Dist: scikit-learn>=1.3; extra == "ml"
+Provides-Extra: dev
+Requires-Dist: pytest>=7.4; extra == "dev"
+Dynamic: license-file
+
+# Vector Engine
+
+ML-first vector computation and retrieval for Python.
+
+Vector Engine provides one clean API for exact search, ANN backends, metadata-aware retrieval, and ML utilities such as kNN and retrieval metrics.
+
+## Why this exists
+
+- ANN libraries are powerful but low-level and backend-specific.
+- Vector DBs solve infra and ops, but many ML workflows need fast local experimentation.
+- Existing ML APIs do not offer a unified, backend-pluggable vector layer for embedding-heavy work.
+
+## Install
+
+```bash
+pip install -e .
+```
+
+Optional extras:
+
+```bash
+pip install -e ".[dev,ml]"
+pip install -e ".[faiss]"
+```
+
+## API contracts (v1.0 target)
+
+- `VectorArray` accepts only 2D arrays with shape `(n, d)` where `n > 0` and `d > 0`.
+- `VectorArray` IDs must be unique and must be `int` or `str`.
+- Metadata length must always align with number of vectors.
+- `VectorIndex.search(..., k=...)` requires `k` to be an integer `> 0`.
+- Score direction is explicit in `Metric.higher_is_better`:
+  - cosine/ip: higher is better
+  - l2: lower is better
+- `kmeans(..., random_state=...)` requires an integer seed and finite vectors.
+- `mine_hard_negatives` supports `top1`, `topk_sample`, and `distance_band` strategies with explicit exclusion controls.
+- Retrieval eval validates malformed ground-truth shapes/types with stable `eval_error` messages.
+
+## 60-second quickstart
+
+```python
+import numpy as np
+from vector_engine import VectorArray, VectorIndex
+
+xb = VectorArray.from_numpy(
+    np.random.randn(1000, 384).astype("float32"),
+    ids=[f"doc-{i}" for i in range(1000)],
+    normalize=True,
+)
+xq = VectorArray.from_numpy(np.random.randn(2, 384).astype("float32"), normalize=True)
+
+index = VectorIndex.create(xb, metric="cosine", backend="bruteforce")
+res = index.search(xq, k=5)
+print(res.ids[0], res.scores[0])
+```
+
+## Core API
+
+- `VectorArray`: canonical vector storage with IDs and metadata.
+- `Metric`: built-in and custom metric definitions.
+- `VectorIndex`: backend-agnostic build/add/search/save/load.
+- `vector_engine.ml`: `knn_classify`, `knn_regress`, `kmeans`.
+- `vector_engine.training`: `mine_hard_negatives` with configurable sampling strategies.
+- `vector_engine.eval`: `precision_at_k`, `recall_at_k`, `ndcg_at_k`, `retrieval_report`, `retrieval_report_detailed`, `batch_metrics_summary`.
+
+## Backend support matrix
+
+| Backend | Search | Add | Save/Load | Custom Metric |
+| --- | ---: | ---: | ---: | ---: |
+| `bruteforce` | yes | yes | yes | yes |
+| `faiss` | yes | yes | yes | no |
+
+## Examples and notebooks
+
+- `notebooks/01_semantic_search.ipynb`
+- `notebooks/02_knn_baseline.ipynb`
+- `notebooks/03_recommender_similarity.ipynb`
+
+## Benchmarks
+
+Run:
+
+```bash
+python benchmarks/compare_bruteforce_vs_faiss.py --mode exact
+```
+
+With exact-overlap gate and artifact output:
+
+```bash
+python benchmarks/compare_bruteforce_vs_faiss.py --mode exact --min-flat-overlap 0.99 --output artifacts/benchmark_exact.json
+```
+
+Benchmark matrix (publishable aggregate):
+
+```bash
+python scripts/benchmark_matrix.py --mode exact --warmup 2 --loops 8 --seed 7 --min-flat-overlap 0.99 --output-dir artifacts/benchmark_matrix
+```
+
+Compose publishable summary bundle:
+
+```bash
+python scripts/publishable_results.py --matrix-summary artifacts/benchmark_matrix/matrix_summary.json --stability-summary artifacts/testing_runs/stability_summary_bruteforce_200.json --output artifacts/benchmark_matrix/publishable_results.v1.json
+```
+
+ANN mode (optional):
+
+```bash
+python benchmarks/compare_bruteforce_vs_faiss.py --mode ann
+```
+
+The benchmark reports:
+
+- `qps`: queries per second (higher is better)
+- `latency_p50_ms` and `latency_p95_ms`: median and tail latency (lower is better)
+- `overlap_vs_bruteforce`: top-k neighbor overlap against exact brute-force (closer to `1.0` is better)
+- `memory_mb_estimate`: coarse in-process memory estimate for vector/query buffers
+
+Recommended protocol for publishable results:
+
+- Use fixed seed and fixed hardware notes.
+- Warm up before timed runs.
+- Run at least 3 repeated trials and report median numbers.
+- Keep dataset size (`n`, `d`, `nq`, `k`) fixed across backend comparisons.
+- Include machine-readable matrix summary artifacts in release evidence.
+
+## Validation snapshot
+
+Artifacts produced in this repo:
+
+- Real-corpus style 3-run reports:
+  - `artifacts/real_corpus_runs/run_1.json`
+  - `artifacts/real_corpus_runs/run_2.json`
+  - `artifacts/real_corpus_runs/run_3.json`
+- Faiss Flat exact-equivalence checks (3 runs):
+  - `artifacts/faiss_equivalence/run_1.json`
+  - `artifacts/faiss_equivalence/run_2.json`
+  - `artifacts/faiss_equivalence/run_3.json`
+- 200-run stability study:
+  - `artifacts/testing_runs/stability_runs_bruteforce_200.jsonl`
+  - `artifacts/testing_runs/stability_summary_bruteforce_200.json`
+  - `artifacts/testing_runs/stability_plot_p95_qps.png`
+- Matrix benchmark summary:
+  - `artifacts/benchmark_matrix/matrix_summary.json`
+  - `artifacts/benchmark_matrix/publishable_results.v1.json`
+
+Observed outcomes for current mock/public-safe corpus:
+
+- 3-run quality is identical across runs (`recall@1/3/6 = 1.0`, `ndcg@1/3/6 = 1.0`).
+- 3-run latency envelope: p95 ranges from `0.0376 ms` to `0.0717 ms`.
+- Faiss Flat exact mode achieves `overlap_vs_bruteforce = 1.0` for all 3 runs with `--min-flat-overlap 0.99`.
+- In exact benchmark runs (`n=10000`, `d=128`, `nq=200`, `k=10`), Faiss Flat p95 latency is `4.17-15.03 ms` vs bruteforce `29.99-37.63 ms`.
+- 200-run bruteforce study: p95 mean `0.0255 ms` (95% interval `0.0203-0.0547 ms`), QPS mean `188,097` (95% interval `117,499-214,111`).
+
+Run the 200-run study:
+
+```bash
+python3 scripts/stability_runs.py \
+  --embeddings artifacts/real_corpus_inputs/embeddings.npy \
+  --query-embeddings artifacts/real_corpus_inputs/query_embeddings.npy \
+  --ids artifacts/real_corpus_inputs/ids.json \
+  --ground-truth artifacts/real_corpus_inputs/ground_truth.json \
+  --metadata artifacts/real_corpus_inputs/metadata.json \
+  --backend bruteforce \
+  --k 6 \
+  --ks 1,3,6 \
+  --loops 5 \
+  --run-count 200 \
+  --output-dir artifacts/testing_runs \
+  --threshold-recall 0.75 \
+  --threshold-ndcg 0.70 \
+  --threshold-p95-ms 120
+```
+
+Example result table format:
+
+| Backend | QPS | p50 ms | p95 ms | overlap@k vs brute-force |
+| --- | ---: | ---: | ---: | ---: |
+| bruteforce | ... | ... | ... | 1.000 |
+| faiss_flat | ... | ... | ... | ... |
+| faiss_ivf (optional) | ... | ... | ... | ... |
+
+## Integration quickstarts
+
+### Local RAG app path
+
+```bash
+pip install -e ".[dev,ml]"
+python examples/minimal_rag_integration.py
+```
+
+### Batch evaluation path
+
+```bash
+python scripts/rag_real_corpus_eval.py \
+  --embeddings artifacts/real_corpus_inputs/embeddings.npy \
+  --query-embeddings artifacts/real_corpus_inputs/query_embeddings.npy \
+  --ids artifacts/real_corpus_inputs/ids.json \
+  --ground-truth artifacts/real_corpus_inputs/ground_truth.json \
+  --metadata artifacts/real_corpus_inputs/metadata.json \
+  --output artifacts/real_corpus_runs/run_1.json \
+  --backend bruteforce \
+  --k 6 \
+  --ks 1,3,6 \
+  --loops 5 \
+  --threshold-recall 0.75 \
+  --threshold-ndcg 0.70 \
+  --threshold-p95-ms 120
+```
+
+### Benchmark interpretation path
+
+```bash
+python benchmarks/compare_bruteforce_vs_faiss.py \
+  --mode exact \
+  --min-flat-overlap 0.99 \
+  --output artifacts/faiss_equivalence/run_1.json
+```
+
+- If `overlap_vs_bruteforce` is near `1.0`, approximation risk is low for that configuration.
+- Use `latency_p95_ms` for user-facing SLO decisions.
+- Use repeated runs + median values before publishing backend comparisons.
+
+### Minimal production path (copy-paste)
+
+```bash
+pip install -e ".[dev,ml]"
+python scripts/rag_baseline.py --output-dir artifacts --k 3
+python scripts/rag_real_corpus_eval.py --embeddings ... --query-embeddings ... --ids ... --ground-truth ... --output artifacts/real_corpus_runs/run_1.json --backend bruteforce --k 10 --ks 1,5,10 --loops 5
+python scripts/stability_runs.py --embeddings ... --query-embeddings ... --ids ... --ground-truth ... --backend bruteforce --run-count 200 --output-dir artifacts/testing_runs
+python benchmarks/compare_bruteforce_vs_faiss.py --mode exact --min-flat-overlap 0.99 --output artifacts/faiss_equivalence/run_1.json
+python scripts/benchmark_matrix.py --mode exact --warmup 2 --loops 8 --seed 7 --min-flat-overlap 0.99 --output-dir artifacts/benchmark_matrix
+python scripts/publishable_results.py --matrix-summary artifacts/benchmark_matrix/matrix_summary.json --stability-summary artifacts/testing_runs/stability_summary_bruteforce_200.json --output artifacts/benchmark_matrix/publishable_results.v1.json
+```
+
+Expected artifacts:
+
+- `artifacts/rag_baseline_metrics.v1.json`
+- `artifacts/real_corpus_runs/run_*.json`
+- `artifacts/testing_runs/stability_summary_*.json`
+- `artifacts/faiss_equivalence/run_*.json`
+- `artifacts/benchmark_matrix/matrix_summary.json`
+- `artifacts/benchmark_matrix/publishable_results.v1.json`
+
+Further reading:
+
+- `docs/integration_guides.md`
+- `docs/reproducibility.md`
+- `docs/kpi_charter.md`
+- `docs/research_claims.md`
+- `docs/credibility_audit.md`
+- `docs/limitations.md`
+- `docs/releases/v1.0.0.md`
+- `docs/paper/reproducibility_appendix.md`
+
+## Publication and release bundle
+
+Generate a release-bundle manifest that checks required docs/governance/evidence files:
+
+```bash
+python scripts/build_release_bundle.py --output-dir artifacts/release_bundle
+```
+
+## Artifact policy (publish vs private)
+
+- Safe to publish:
+  - benchmark result summaries
+  - stability aggregate summaries
+  - synthetic/mock input examples
+- Keep private:
+  - real corpus raw embeddings
+  - query embeddings derived from private data
+  - sensitive metadata and ID mappings
+- Recommended:
+  - commit docs + summary metrics in repo
+  - keep private input blobs in external storage
+
+## Project adoption checklist
+
+- Install: `pip install -e ".[dev,ml]"` and optional `.[faiss]`.
+- Validation: run `pytest -q`.
+- Quality baseline: run `python scripts/rag_baseline.py`.
+- Real corpus eval: run `python scripts/rag_real_corpus_eval.py --embeddings ... --query-embeddings ... --ids ... --ground-truth ... --threshold-recall 0.75 --threshold-ndcg 0.70 --threshold-p95-ms 120`.
+- Persistence: verify `VectorIndex.save/load` on your own embeddings snapshot.
+- Performance: run benchmark script with your target `n`, `d`, `nq`, `k`.
+- Integration: run `python examples/minimal_rag_integration.py`.
+
+## Feature snapshot
+
+- `kmeans` returns rich outputs (`labels`, `centers`, `inertia`, `n_iter`) with deterministic validation.
+- Hard-negative mining supports `top1`, `topk_sample`, and `distance_band`, plus `exclude_ids` / `exclude_mask`.
+- Retrieval evaluation includes `retrieval_report_detailed(include_per_query=...)` and `batch_metrics_summary(include_std=True)`.
+- Public demo bootstrap is available under `demo_repo_template/`.
+
+## v1.0 readiness gates
+
+- Benchmark matrix artifacts produced with fixed protocol and environment metadata.
+- Stability harness demonstrates repeatability for latency/QPS/quality summaries.
+- API stability contract documented in `docs/api.md` and enforced in `tests/test_api_stability.py`.
+- Release packaging includes reproducible command blocks and artifact policy.
+
+## Governance and trust
+
+- `LICENSE`
+- `CITATION.cff`
+
+## Error cases
+
+Stable error prefixes are used for fast debugging:
+
+- `vector_array_error`: malformed array, IDs, metadata, subset lookup
+- `metric_error`: unsupported or invalid metric definitions
+- `index_error`: index lifecycle/search/add/persistence consistency issues
+- `manifest_error`: missing/unsupported manifest fields or version
+
+## Troubleshooting
+
+- **Faiss not available**
+  - Install with `pip install -e ".[faiss]"`.
+- **Dimension mismatch at search/add**
+  - Ensure both base vectors and query vectors use the same embedding dimension.
+- **Metric confusion**
+  - For cosine similarity, pass normalized vectors or set `normalize=True`.
+- **Persistence load failure**
+  - Check manifest version compatibility and whether artifacts were modified after save.

vector_engine-1.0.0.dist-info/RECORD

@@ -0,0 +1,24 @@
+vector_engine/__init__.py,sha256=VbXKWPz-SPssPhVvSwzMQvpkaiBRS7K6jEigfyLRWaI,285
+vector_engine/array.py,sha256=mEulAl6fJFFAnB-k2Lav8E93Rw4oIBHsjYioWCiBwWE,5703
+vector_engine/index.py,sha256=wAhkRbzqu9odUbSeF3kmNJZ3o5oOvGrWyqmc3VxO0Sc,7702
+vector_engine/metric.py,sha256=FiYhbOrUuX9u84EeCgjrrg4np-eXKjP5jgILeiNm-Bs,1824
+vector_engine/results.py,sha256=dQ8B6vXQvsR9moGGcbqMc5fnsguN5aDKayQE3fI9Ypc,310
+vector_engine/backends/__init__.py,sha256=-vEe47Gsu-xLR7rv5-Y-DXpcFfT49qHoZR2cdt7YZyM,328
+vector_engine/backends/base.py,sha256=UfY7I3gvbEcEraYeMkd224BNOu_Ces0G7H1THsV7XgI,575
+vector_engine/backends/bruteforce.py,sha256=bOjoFeu__lbswaps2Tnxa5-u4gHTKMPYZ0GvRlkNKWE,4137
+vector_engine/backends/faiss_backend.py,sha256=7x6UT6EEwMZmos_Nh3Q6kiOeS1n3R-WAt2o-OELReZs,4520
+vector_engine/backends/registry.py,sha256=Dzgi4mdztuMz0sM2G_uGkKB-FNy5ZVk8VHthLZmTSiw,347
+vector_engine/eval/__init__.py,sha256=-EPSOvNxi3kaFocdVzXY10ZtCkaI-KKaBNGeiJ5QeH4,318
+vector_engine/eval/retrieval.py,sha256=Ij3DjtE8dhpjgMyXpBQBXOae0GJNr1wzlfMCXo804t8,6586
+vector_engine/io/__init__.py,sha256=ygJY89J9bOexA4lU2A9o1cC_PkeyR0IzD58_vrFBx6w,129
+vector_engine/io/manifest.py,sha256=cIsF5QtwuUJGmsNGIOLcQqyBSciREEUgswQClD6MGLY,1385
+vector_engine/ml/__init__.py,sha256=NmB0JvD7Y8UiaYnM6vXA45GPBosmKNSZAzY80NgFMnc,157
+vector_engine/ml/clustering.py,sha256=clz52N7_-A08roL5k5_U4ZqKCQe5R7lThX6_dIGNq5M,1664
+vector_engine/ml/knn.py,sha256=2Rk8hBgKvpimLJKjDuCv33n86-00MEh_AF0WZK-pZ_Q,2315
+vector_engine/training/__init__.py,sha256=XJ-EQTrFwA7TnWEfPgvkx9KoN_FVnX0pspfG1vCOUhE,112
+vector_engine/training/hard_negative.py,sha256=_ZnvwD0XbNcPUvvCvW71VMJkvlSH5wv7pPPCNX-gKbw,4873
+vector_engine-1.0.0.dist-info/licenses/LICENSE,sha256=Xy3nx0VJ3Dd9vtPu7o1CNZ0sC1jHf1jdNQlQ_qY8mp0,1083
+vector_engine-1.0.0.dist-info/METADATA,sha256=t5HIUlDOpGJaUfWnxoe8iLAM9dHVGgZ_QYJ5Yx_7EW8,12583
+vector_engine-1.0.0.dist-info/WHEEL,sha256=YCfwYGOYMi5Jhw2fU4yNgwErybb2IX5PEwBKV4ZbdBo,91
+vector_engine-1.0.0.dist-info/top_level.txt,sha256=DWhv4njcslmKqCi_n2IiOUrrYMwEV5IvXiWpkaQAf5A,14
+vector_engine-1.0.0.dist-info/RECORD,,

vector_engine-1.0.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (82.0.0)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

vector_engine-1.0.0.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Vector Engine Contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

vector_engine-1.0.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+vector_engine