lucivy

BM25 search engine with cross-token fuzzy matching — it finds substrings, handles typos, and matches across word boundaries. Built for code search, technical docs, and as a BM25 complement to vector databases.

Install

Everything is MIT-licensed.

Quick start

import lucivy

index = lucivy.Index.create("./my_index", fields=[
    {"name": "title", "type": "text"},
    {"name": "body", "type": "text"},
])

index.add(1, title="Rust Programming", body="Systems programming with memory safety")
index.add(2, title="Python Guide", body="Data science and web development")
index.commit()

results = index.search("programming", highlights=True)
for r in results:
    print(r.doc_id, r.score, r.highlights)

See the language-specific READMEs for full API docs:

• Python
• Node.js
• WASM / browser

Query types

lucivy queries operate on stored text (cross-token). They handle multi-word phrases, substrings, separators, and special characters naturally.

contains — the workhorse query

Fuzzy substring match with separator awareness.

# Exact substring
index.search({"type": "contains", "field": "body", "value": "programming language"})

# Substring within a token: "program" matches "programming"
index.search({"type": "contains", "field": "body", "value": "program"})

# Fuzzy tolerance (default distance=1, catches typos)
index.search({"type": "contains", "field": "body", "value": "programing languag", "distance": 1})

# Strict exact: distance=0 disables fuzzy
index.search({"type": "contains", "field": "body", "value": "programming", "distance": 0})

contains + regex

Regex on stored text (cross-token).

# Matches "programming language" — the .* spans the space between tokens
index.search({"type": "contains", "field": "body", "value": "program.*language", "regex": True})

# Alternation
index.search({"type": "contains", "field": "body", "value": "python|rust", "regex": True})

contains_split

Splits query into words, each word is a contains, combined with OR.

# String query (auto contains_split across all text fields)
index.search("rust async programming")

# Explicit dict query on a specific field
index.search({"type": "contains_split", "field": "body", "value": "memory safety"})

boolean

Combine sub-queries with must (AND), should (OR), must_not (NOT).

index.search({
    "type": "boolean",
    "must": [
        {"type": "contains", "field": "body", "value": "rust"},
        {"type": "contains", "field": "body", "value": "programming"},
    ],
    "must_not": [{"type": "contains", "field": "body", "value": "javascript"}],
})

Filters on non-text fields

Non-text fields (i64, f64, u64, keyword) can be filtered via the filters key.

index.search({
    "type": "contains",
    "field": "body",
    "value": "programming",
    "filters": [
        {"field": "year", "op": "gte", "value": 2023},
    ],
})
# Supported ops: eq, ne, lt, lte, gt, gte, in, not_in, between, starts_with, contains

Highlights

All query types support byte-offset highlights. Internal fields (._raw, ._ngram) are automatically filtered out.

results = index.search("rust programming", highlights=True)
for r in results:
    if r.highlights:
        for field, offsets in r.highlights.items():
            print(f"  {field}: {offsets}")  # e.g. "body": [(5, 9), (20, 31)]

Fields (stored values)

Retrieve stored field values alongside search results — useful for displaying file names, titles, or content excerpts.

results = index.search("rust programming", fields=True)
for r in results:
    print(r.doc_id, r.score, r.fields['title'])

Snapshots (export / import)

Export an index to a portable .luce binary blob, import it elsewhere.

index.export_snapshot_to("./backup.luce")
restored = lucivy.Index.import_snapshot_from("./backup.luce", dest_path="./restored")

What contains matches

Fuzzy mode (default):

Regex mode (regex: true):

Internals

Triple-field layout

Every text field automatically gets 3 sub-fields:

This is transparent to the user — you always reference the base field name.

NgramContainsQuery — how contains works

1. Candidate collection — depends on mode:
   • Fuzzy: term dictionary lookup on ._raw (O(1) via FST), falling back to trigram intersection on ._ngram if the exact term isn't found
   • Regex: trigram union on ._ngram from extracted regex literals
   • Short literals: full segment scan when literals < 3 chars
2. Verification — read stored text, dispatch to fuzzy or regex verifier
3. BM25 scoring — standard idf * (1 + k1) * tf / (tf + k1 * (1 - b + b * dl / avgdl))

Building from source

# Rust library tests
cargo test --lib

# Python bindings
cd bindings/python
maturin develop --release
pytest tests/ -v

# Node.js bindings
cd bindings/nodejs && npm run build
node test.mjs

# C++ bindings
cargo build -p lucivy-cpp --release

Lineage

Fork of tantivy v0.26.0 (via izihawa/tantivy).

quickwit-oss/tantivy v0.22
  -> izihawa/tantivy v0.26.0 (regex phrase queries, FST improvements)
    -> L-Defraiteur/lucivy (NgramContainsQuery, contains_split, fuzzy/regex/hybrid modes, HighlightSink, Python/Node.js/C++/WASM bindings)

License

MIT. See LICENSE.

Fork of tantivy v0.26.0, also MIT (see NOTICE).