lodedb 1.3.0__tar.gz → 1.3.2__tar.gz

{lodedb-1.3.0 → lodedb-1.3.2}/Cargo.toml

@@ -7,7 +7,7 @@ members = [
 resolver = "2"
 
 [workspace.package]
-version = "1.3.0"
+version = "1.3.2"
 edition = "2021"
 rust-version = "1.70"
 license = "Apache-2.0"

{lodedb-1.3.0 → lodedb-1.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lodedb
-Version: 1.3.0
+Version: 1.3.2
 Classifier: Development Status :: 4 - Beta
 Classifier: Intended Audience :: Developers
 Classifier: Programming Language :: Python :: 3 :: Only
@@ -10,6 +10,7 @@ Requires-Dist: numpy>=2.0.0,<3
 Requires-Dist: typer>=0.12.0
 Requires-Dist: pyyaml>=6.0.0
 Requires-Dist: lodedb[embeddings,torch,image,mcp,langchain,llama-index,mem0] ; extra == 'all'
+Requires-Dist: cognee>=1.1.0,<2 ; python_full_version < '3.15' and extra == 'cognee'
 Requires-Dist: pytest>=8.3.0 ; extra == 'dev'
 Requires-Dist: ruff==0.15.17 ; extra == 'dev'
 Requires-Dist: onnxruntime>=1.20.0,<2 ; extra == 'embeddings'
@@ -24,6 +25,7 @@ Requires-Dist: mem0ai>=2.0.0 ; extra == 'mem0'
 Requires-Dist: optimum[exporters]>=1.23.0,<2.0.0 ; extra == 'onnx-export'
 Requires-Dist: sentence-transformers>=3.0.0,<5 ; extra == 'torch'
 Provides-Extra: all
+Provides-Extra: cognee
 Provides-Extra: dev
 Provides-Extra: embeddings
 Provides-Extra: gpu
@@ -115,7 +117,8 @@ stays **sub-millisecond even at 1M vectors**.
   [How it works](#multimodal--bring-your-own-vectors).
 - **Batteries included**: a `lodedb` CLI, a loopback/private-network dev server, an
   [MCP server](#use-as-an-mcp-server), LangChain, LlamaIndex, and mem0 adapters
-  (`VectorStore`s, plus a LlamaIndex `PropertyGraphStore`), and a one-line
+  (`VectorStore`s, plus a LlamaIndex `PropertyGraphStore`), a
+  [cognee](https://github.com/topoteretes/cognee) vector-DB provider, and a one-line
   [PrivateGPT](https://github.com/zylon-ai/private-gpt) vector-store provider built on the
   LlamaIndex adapter.
 - **Swift / iOS bindings**: a native Swift package for macOS and iOS over the same Rust
@@ -156,8 +159,9 @@ pip install "lodedb[embeddings,torch]"               # + PyTorch fallback, CLIP,
 pip install "lodedb[gpu]"                            # GPU-resident scan (Linux/CUDA)
 pip install "lodedb[image]"                          # image + text (CLIP) embedding (model="clip")
 pip install "lodedb[mcp,langchain,llama-index,mem0]" # MCP server + LangChain/LlamaIndex/mem0 adapters
+pip install "lodedb[cognee]"                         # register LodeDB as a cognee vector_db_provider
 pip install "lodedb[onnx-export]"                    # export ONNX for a custom model (Optimum); presets need no export
-pip install "lodedb[all]"                            # everything above
+pip install "lodedb[all]"                            # everything above (except gpu/onnx-export/cognee)
 ```
 
 Using LodeDB as memory for a coding assistant? After installing the `mcp,embeddings` extras,
@@ -384,6 +388,42 @@ db.search("turbine fault", k=5)                                 # nearest cluste
 `ann_clusters` (partitions) defaults to about `sqrt(n)` and `ann_nprobe` (clusters probed per
 query) to about `sqrt(clusters)`. ANN is a create-time choice persisted with the index and works
 for text and bring-your-own-vector indexes alike; keep the exact default for small-to-mid corpora.
+On reopen, `ann_nprobe` is a session-only override, so benchmark handles can change probe breadth
+without rebuilding; the algorithm and partition count remain the persisted configuration.
+
+For large corpora, these are useful starting ranges to measure, not recall or latency promises:
+
+| Knob | Useful range | Trade-off |
+| --- | --- | --- |
+| `ann_nprobe` | 64 to 256 clusters | More probes raise recall and scan work; probing every cluster is exact. |
+| `rescore_oversample` | 2 to 8 times `k` | More candidates improve the chance that fp32 re-ranking repairs compact-code errors, at more sidecar reads. |
+</details>
+
+<details>
+<summary><b>Two-stage rescore (`rescore="original"`)</b></summary>
+
+TurboVec's compact 4-bit scan is fast, but its quantized first-stage ranking can put close
+neighbors in the wrong order. Enable rescore at creation to retain each input vector in a separate
+original-precision sidecar, then re-score the first-stage candidate pool with fp32 dot products:
+
+```python
+db = LodeDB(
+    path="./data",
+    ann="cluster",
+    ann_clusters=1024,
+    rescore="original",
+    rescore_dtype="float16",      # default; about 2 bytes per dimension per vector
+    rescore_oversample=4,
+)
+```
+
+`float16` sidecars add about 2 bytes per dimension per vector (`float32` adds 4; `int8` adds 1).
+They let the final ordering rise above the practical 4-bit ranking ceiling while retaining the
+compact index for candidate generation. Re-scored candidate scores are exact fp32 dots, but the
+overall result can still be approximate when ANN or a too-small candidate pool excludes a true
+neighbor. Rescore is a create-time capture choice: originals are not recoverable after a normal
+store's first ingest, so enabling it later requires a rebuild. On a rescore-enabled reopen,
+`rescore_oversample` is a session-only override; mode and dtype must match the persisted sidecar.
 </details>
 
 ## Multimodal & bring-your-own vectors
@@ -533,6 +573,9 @@ lodedb mcp         # stdio MCP server for agent memory
 lodedb benchmark   # local, metrics-only benchmark
 ```
 
+Use `lodedb serve` as a local OpenAI-compatible embeddings provider; see [LodeDB as a local
+embeddings provider](docs/integrations.md#lodedb-as-a-local-embeddings-provider).
+
 ## Use as an MCP server
 
 LodeDB ships a [Model Context Protocol](https://modelcontextprotocol.io) server, so an agent

{lodedb-1.3.0 → lodedb-1.3.2}/README.md

@@ -66,7 +66,8 @@ stays **sub-millisecond even at 1M vectors**.
   [How it works](#multimodal--bring-your-own-vectors).
 - **Batteries included**: a `lodedb` CLI, a loopback/private-network dev server, an
   [MCP server](#use-as-an-mcp-server), LangChain, LlamaIndex, and mem0 adapters
-  (`VectorStore`s, plus a LlamaIndex `PropertyGraphStore`), and a one-line
+  (`VectorStore`s, plus a LlamaIndex `PropertyGraphStore`), a
+  [cognee](https://github.com/topoteretes/cognee) vector-DB provider, and a one-line
   [PrivateGPT](https://github.com/zylon-ai/private-gpt) vector-store provider built on the
   LlamaIndex adapter.
 - **Swift / iOS bindings**: a native Swift package for macOS and iOS over the same Rust
@@ -107,8 +108,9 @@ pip install "lodedb[embeddings,torch]"               # + PyTorch fallback, CLIP,
 pip install "lodedb[gpu]"                            # GPU-resident scan (Linux/CUDA)
 pip install "lodedb[image]"                          # image + text (CLIP) embedding (model="clip")
 pip install "lodedb[mcp,langchain,llama-index,mem0]" # MCP server + LangChain/LlamaIndex/mem0 adapters
+pip install "lodedb[cognee]"                         # register LodeDB as a cognee vector_db_provider
 pip install "lodedb[onnx-export]"                    # export ONNX for a custom model (Optimum); presets need no export
-pip install "lodedb[all]"                            # everything above
+pip install "lodedb[all]"                            # everything above (except gpu/onnx-export/cognee)
 ```
 
 Using LodeDB as memory for a coding assistant? After installing the `mcp,embeddings` extras,
@@ -335,6 +337,42 @@ db.search("turbine fault", k=5)                                 # nearest cluste
 `ann_clusters` (partitions) defaults to about `sqrt(n)` and `ann_nprobe` (clusters probed per
 query) to about `sqrt(clusters)`. ANN is a create-time choice persisted with the index and works
 for text and bring-your-own-vector indexes alike; keep the exact default for small-to-mid corpora.
+On reopen, `ann_nprobe` is a session-only override, so benchmark handles can change probe breadth
+without rebuilding; the algorithm and partition count remain the persisted configuration.
+
+For large corpora, these are useful starting ranges to measure, not recall or latency promises:
+
+| Knob | Useful range | Trade-off |
+| --- | --- | --- |
+| `ann_nprobe` | 64 to 256 clusters | More probes raise recall and scan work; probing every cluster is exact. |
+| `rescore_oversample` | 2 to 8 times `k` | More candidates improve the chance that fp32 re-ranking repairs compact-code errors, at more sidecar reads. |
+</details>
+
+<details>
+<summary><b>Two-stage rescore (`rescore="original"`)</b></summary>
+
+TurboVec's compact 4-bit scan is fast, but its quantized first-stage ranking can put close
+neighbors in the wrong order. Enable rescore at creation to retain each input vector in a separate
+original-precision sidecar, then re-score the first-stage candidate pool with fp32 dot products:
+
+```python
+db = LodeDB(
+    path="./data",
+    ann="cluster",
+    ann_clusters=1024,
+    rescore="original",
+    rescore_dtype="float16",      # default; about 2 bytes per dimension per vector
+    rescore_oversample=4,
+)
+```
+
+`float16` sidecars add about 2 bytes per dimension per vector (`float32` adds 4; `int8` adds 1).
+They let the final ordering rise above the practical 4-bit ranking ceiling while retaining the
+compact index for candidate generation. Re-scored candidate scores are exact fp32 dots, but the
+overall result can still be approximate when ANN or a too-small candidate pool excludes a true
+neighbor. Rescore is a create-time capture choice: originals are not recoverable after a normal
+store's first ingest, so enabling it later requires a rebuild. On a rescore-enabled reopen,
+`rescore_oversample` is a session-only override; mode and dtype must match the persisted sidecar.
 </details>
 
 ## Multimodal & bring-your-own vectors
@@ -484,6 +522,9 @@ lodedb mcp         # stdio MCP server for agent memory
 lodedb benchmark   # local, metrics-only benchmark
 ```
 
+Use `lodedb serve` as a local OpenAI-compatible embeddings provider; see [LodeDB as a local
+embeddings provider](docs/integrations.md#lodedb-as-a-local-embeddings-provider).
+
 ## Use as an MCP server
 
 LodeDB ships a [Model Context Protocol](https://modelcontextprotocol.io) server, so an agent

{lodedb-1.3.0 → lodedb-1.3.2}/crates/lodedb-core/examples/write_generation_fixture.rs

@@ -57,6 +57,7 @@ fn main() {
             lexical_tokens: None,
             multivec: None,
             ann: None,
+            tvvf_manifest: None,
             compress_text: true,
         },
         GenerationWriteOptions::default(),

{lodedb-1.3.0 → lodedb-1.3.2}/crates/lodedb-core/examples/write_wal_fixture.rs

@@ -58,6 +58,7 @@ fn main() {
             lexical_tokens: None,
             multivec: None,
             ann: None,
+            tvvf_manifest: None,
             compress_text: true,
         },
         GenerationWriteOptions::default(),