warm-rust 0.6.5

Local-first search engine and agent memory: fast approximate nearest-neighbor and BM25 full-text search, encrypted stores, provable deletion, built on vector-symbolic architectures.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114

# Copyright 2024-2026 WritersLogic Contributors
# SPDX-License-Identifier: Apache-2.0
#
# The single source for the capability copy on the landing page and in the footer. Other work
# updates this file when a runtime, a feature or a measured result lands; nothing on the landing
# page names a runtime or feature that is not listed here.
#
# Contract, enforced by scripts/docs_site.py (a violation fails the strict docs build):
#   runtimes[].status   "available" (shipped, install tab on the landing page) or "preview"
#                       (shown as a chip, no tab). Planned work is not listed at all.
#   runtimes[].requires paths that must exist in the repository for an "available" entry.
#   runtimes[].snippets code sections included from the executed quickstart files, written
#                       "path:section". They are read at build time, so they cannot drift.
#   headline            names a committed results file and the dotted path of a metric in it.
#   comparison          optional docs/_data/comparison.json written by the benchmark generator
#                       fills the comparison table; without it the page derives a table from
#                       the committed results named under comparison_fallback.

technical_line: Hybrid search, graph indexing, and vector-symbolic memory.
tagline: Search and memory that run entirely on your machine.
lede: >-
  WARM is a local-first document search engine and a vector-symbolic associative memory in one
  store. Hybrid BM25 and exact-cosine retrieval, byte-offset hits, crash-safe persistence, and
  superposed memory you can bind, bundle and query. No service calls and no telemetry.
license: Apache-2.0

# The announcement bar. Leave `text` empty to hide it.
announcement:
  text: ""
  url: ""

runtimes:
  - id: rust
    name: Rust
    status: available
    install: cargo add warm-rust --features security
    requires: [Cargo.toml, examples/quickstart/rust/src/main.rs]
    snippets:
      - examples/quickstart/rust/src/main.rs:ingest
      - examples/quickstart/rust/src/main.rs:hit
    note: The engine itself. Every other runtime is a surface over it.
  - id: python
    name: Python
    status: available
    install: uv pip install warm-sdk
    requires: [bindings/python/warm-sdk/pyproject.toml, examples/quickstart/python/quickstart.py]
    snippets:
      - examples/quickstart/python/quickstart.py:ingest
      - examples/quickstart/python/quickstart.py:hit
    note: An HTTP client for warm-server. The warm-python wheel provides native vector search and phase vectors.
  - id: node
    name: Node
    status: available
    install: npm install writerslogic-warm
    requires: [index.d.ts, examples/quickstart/node/quickstart.mjs]
    snippets:
      - examples/quickstart/node/quickstart.mjs:ingest
      - examples/quickstart/node/quickstart.mjs:hit
    note: The npm package installs a native binary for your platform. Building from source requires Rust.

features:
  - title: Hybrid search
    body: Stemmed BM25 fused with exact cosine over embeddings you supply. Hits carry the passage and its byte range.
    link: quickstart/index.md
  - title: Vector-symbolic memory
    body: Sparse binary hypervectors with bind, bundle and permute, a triple store with multi-hop queries, Hopfield cleanup and a resonator network.
    link: concepts/holographic.md
  - title: Crash-safe storage
    body: Atomic batches, a mmap segmented log with generations, and replay. A store is a directory.
    link: production-readiness.md
  - title: Hardening you opt into
    body: AES-256-GCM stores, signed audit logs, credential-gated agents and COSE/SCITT provenance behind Cargo features.
    link: PRIVACY.md
  - title: Local by construction
    body: The engine makes no network calls and sends no telemetry. The optional HTTP server is yours to run.
    link: concepts/privacy-limits.md
  - title: Measured, including losses
    body: Every number on this site traces to a committed results file, and the benchmark pages say where another system wins.
    link: PUBLIC-BENCHMARKS.md

# The headline result. `file` is relative to the repository root; each metric is a dotted path.
headline:
  title: Hybrid retrieval on SciFact
  file: benchmarks/results/public_scifact.json
  metric: metrics.hms_hybrid.ndcg_at_10
  metric_label: nDCG@10
  compare:
    - label: WARM dense (exact cosine)
      metric: metrics.hms_dense_exact.ndcg_at_10
    - label: WARM lexical (stemmed BM25)
      metric: metrics.hms_lexical.ndcg_at_10
  context: BEIR SciFact test split, 300 queries over 5,183 documents, MiniLM-L6-v2 embeddings (384 dimensions).
  caveat: The published BM25 figure (0.665, Thakur et al. 2021) is quoted from the BEIR paper, not re-run here.
  published:
    - label: BM25 as published in BEIR
      value: 0.665
      source: https://arxiv.org/abs/2104.08663
  link: PUBLIC-BENCHMARKS.md

# Vector-search chart: queries per second at each recall target, single thread, one query at a
# time. Systems that do not reach a target are listed as such, never dropped.
chart:
  title: Vector search on nytimes-256-angular
  file: benchmarks/results/public_qgraph_nytimes-256-angular.json
  recall_targets: ["0.90", "0.95"]
  warm_prefix: HMS
  # Rows the benchmark page itself marks as unconfirmed. They stay on the chart, marked, and are
  # left out of the "fastest WARM configuration" sentence.
  annotations:
    "WARM VGraph (4-bit codes, rerank)": opt-in encoding that ran at lower machine load than the other rows; the benchmark page treats it as unconfirmed.
  link: PUBLIC-BENCHMARKS.md

comparison_fallback:
  - benchmarks/results/public_scifact.json