Skip to content

Code graph and memory for your coding agents, in one Rust program. For Claude Code, Codex, Cursor and any MCP client.

13.2 KB
Raw Blame

Graphicle graph engine: third-party notices

Copyright (c) 2026 KarLex AI, Inc. The graph engine is licensed under the Apache License, Version 2.0 (LICENSE); NOTICE carries the notices that Apache-2.0 section 4(d) requires. This file describes each adapted source and where its licence text is kept.

This package ports Graphify 0.9.95, commit 92741d162b2094ffc2fed58cfbd8ee823e9da62c. Original authors and notices are retained in licenses/. Adapted implementation files identify the corresponding upstream file.

The original application code is translated into Rust. The migration ledger distinguishes runtime ports, tests, unchanged fixtures, assets and build tooling. Original Graphify files remain in the pinned source archive and extracted tree. No completeness or performance claim follows from copying an interface.

Unicode case-folding, word-character and identifier tables are generated from CPython's Unicode 15.0.0 database for compatibility with the pinned Python 3.12 reference runtime. NFKC normalization uses the separately licensed unicode-normalization crate. The Unicode data license is retained in licenses/UNICODE.txt.

The Markdown port's YAML-1.1 scalar grammars follow PyYAML; its MIT notice is retained in licenses/PyYAML-LICENSE. YAML parsing itself uses yaml-rust2.

HTML5 character-reference data and unescape rules are adapted from CPython 3.12 into native Rust; the PSF license and copyright notice are retained in licenses/PYTHON-LICENSE.txt. No Python interpreter is linked or invoked.

Summary of changes to CPython 3.12 (PSF License Version 2, section 3). Each part below is translated into Rust, limited to the behaviour the pinned Python reference relies on, and has no Python runtime: html.unescape and its HTML5 entity table (src/html_compat.rs, src/html_entities.json); datetime.fromisoformat (src/python_datetime.rs); textwrap.shorten and TextWrapper._split (src/extract_languages.rs); the str hash (SipHash-1-3) and the iteration order of string sets (src/python_compat.rs); random.Random seeding, shuffle and sample (src/cluster.rs); the Unicode 15.0.0 case-folding, word-character and printable tables (src/unicode15.rs, src/python_repr_data.rs, src/regex_compat.rs); and the ipaddress reserved-network tables (src/ip_policy_data.rs).

The seeded generators reproduce the reference MT19937 algorithm of Makoto Matsumoto and Takuji Nishimura, as CPython's random module and NumPy's RandomState seed it (src/cluster.rs, src/minhash.rs); its BSD-3-Clause notice is in licenses/MT19937-LICENSE.txt. The string hash follows the SipHash implementation in CPython's Python/pyhash.c by Marek Majkowski (MIT); its notice is in licenses/SipHash-csiphash-LICENSE.txt.

The Terraform parser retains the exact generated C grammar from the pinned tree-sitter-hcl 1.2.0 source distribution, with a native Rust binding. Its Apache-2.0 license and checksummed provenance are in vendor/tree-sitter-hcl/. This is a parser dependency, not an additional Graphify application port.

The SQL extractor similarly retains the exact generated C grammar from tree-sitter-sql 0.3.11, with a native Rust binding. Its MIT license and checksummed provenance are in vendor/tree-sitter-sql/.

The R, Solidity, VB.NET and Erlang extractors likewise retain the exact generated C grammars behind Graphify's optional Python parsers, each with a native Rust binding: tree-sitter-vb-dotnet 0.3.0 (MIT), tree-sitter-solidity 1.2.13 (MIT), and the R (r-lib/tree-sitter-r, MIT) and Erlang (WhatsApp/tree-sitter-erlang, Apache-2.0) parsers bundled in tree-sitter-language-pack 0.11.0. Licenses and checksummed provenance are in vendor/tree-sitter-{vb-dotnet,solidity,r,erlang}/; tools/vendor_extractor_grammars.py reproduces them from the pinned archives.

The Python parser retains the exact generated C grammar of tree-sitter-python 0.25.0 (MIT), the version behind Graphify's pinned Python parser, with a native Rust binding. Its external scanner carries one fix from tree-sitter-language-pack 1.20.0 (MIT): the state serializer no longer writes one byte past its 1,024-byte buffer on a deeply indented file. The license, the patch and checksummed provenance are in vendor/tree-sitter-python/; the pack's MIT licence, which covers the patch (Copyright 2025-2026 Kreuzberg, Inc.), is in licenses/vendor/tree-sitter-python__patch__tree-sitter-language-pack-1.20.0__LICENSE.txt.

The Kotlin and PHP parsers likewise retain the exact generated C grammars of tree-sitter-kotlin 1.1.0 (MIT, Amaan Qureshi; its public ABI headers come from the MIT-licensed tree-sitter-java 0.23.5) and tree-sitter-php 0.24.1 (MIT), each with a native Rust binding; licenses and checksummed provenance are in vendor/tree-sitter-{kotlin,php}/. tools/vendor_engine_grammars.py reproduces them and Groovy from the pinned archives.

The Groovy parser retains the generated C grammar of tree-sitter-groovy 0.1.2 (tag v0.1.2 of amaanq/tree-sitter-groovy). Its licence basis is a declaration: the upstream Cargo.toml, retained as vendor/tree-sitter-groovy/UPSTREAM-LICENSE-DECLARATION.toml, names the author Amaan Qureshi and license = "MIT", but neither the tag nor its Python source distribution has a licence file or a copyright notice. We rely on that declaration and ship the standard MIT text with it, without inventing a copyright line (licenses/vendor/tree-sitter-groovy__DECLARED-MIT.txt). On 2026-10-13, v0.1.2 was still the latest tag and the default branch still had no licence file; a later release with one replaces this basis if its parser is identical.

Each vendored grammar crate wraps the upstream generated sources with a small first-party build glue: its Cargo.toml, build.rs and src/lib.rs. That glue is Copyright (c) 2026 KarLex AI, Inc. and is licensed under the same terms as the grammar it wraps, so each crate's license field covers the whole crate. Copies of every vendored licence text are also in licenses/vendor/, so that release archives carry them.

Community detection links the existing MIT-licensed network_partitions Rust kernel under vendor/network_partitions/. Its algorithms match the kernel shipped in graspologic-native 1.2.5 (Copyright (c) Microsoft Corporation), which took the MIT-licensed CWTS networkanalysis Java implementation as its starting point; the CWTS notice is in licenses/vendor/network_partitions__CWTS-networkanalysis-LICENSE.txt. The sources are formatted with rustfmt (seven files differ from the recorded upstream hashes only in formatting), and three explicit borrowed lifetime annotations were added in two files for current Rust diagnostics; PROVENANCE.json retains the upstream hashes and records the two annotated files. The Louvain fallback follows NetworkX 3.7 (BSD-3-Clause), and its seeded shuffle follows CPython 3.12 (PSF). Dependency licenses are retained.

Pascal uses the MIT-licensed grammar from wpostma/tree-sitter-pascal, commit a55068ee73ec993e06d81ca4eb53b85452e79c72, retained under vendor/tree-sitter-pascal/. Graphify's optional Python parser 0.11.0 switched to AGPL-3.0-only. That fork is a development comparison reference only; it is not included in this native package. Both the original Graphify traversal and its regex fallback are translated. Grammar extensions unique to the AGPL fork are outside this dependency compatibility claim; see the Pascal receipt.

Optional PostgreSQL introspection uses the pq-sys Rust bindings and the PostgreSQL libpq client supplied by pq-src. Their upstream licenses and PostgreSQL copyright notice are retained under licenses/. The native client replaces the Python psycopg dependency; it does not embed a Python interpreter.

Document discovery uses the MIT-licensed zip, calamine, and pdf-extract Rust libraries in place of Python ZIP/Office/PDF readers. encoding_rs retains its Apache-2.0/MIT and WHATWG BSD-3-Clause notices; md-5 is MIT/Apache-2.0 and is used only for the original manifest's change-detection digests. Available dependency notices are retained under licenses/; pinned dependency versions are recorded in Cargo.lock. PDF layout whitespace can differ between the native parser and pypdf, as recorded in the discovery comparison receipt.

Graph database export uses the MIT-licensed neo4rs Rust driver (pinned 0.9.0-rc.10, with routing support) and the BSD-3-Clause redis Rust client. The FalkorDB parameter header encoding follows the MIT-licensed FalkorDB Python SDK 1.7.1; its license is retained. No FalkorDB server implementation is bundled. neo4rs's distributed Cargo license/author declaration is retained because its source package and tagged repository contain no standalone license file. Provider transport qualification and its limits are recorded separately.

SVG layout follows NetworkX 3.7's Fruchterman–Reingold force and energy formulations. The large-graph optimizer is the BSD-3-Clause lbfgsb-rs-pure 0.1.2 Rust implementation; its license is retained. Native SVG primitives replace Matplotlib; identical font metrics and large-graph optimizer iterates are not claimed. Unicode digit ranges used by Obsidian are from the same CPython Unicode 15.0.0 reference database described above.

The native Robot Framework source lexer follows the Apache-2.0-licensed Robot Framework 7.5 tokenizer, lexer settings and statement classification. Its built-in language header/setting tables are retained as JSON data. Copyright 2008–2015 Nokia Networks; copyright 2016– Robot Framework Foundation. The license is retained at licenses/robotframework-7.5-LICENSE.txt. The source runtime uses Rust; the Python package is used only as a development oracle.

Model token estimates use the MIT-licensed tiktoken-rs 0.12.0 implementation and its bundled cl100k vocabulary. Its license is retained under licenses/. Bedrock Converse uses the Apache-2.0 AWS SDK for Rust (aws-config 1.12.0 and aws-sdk-bedrockruntime 1.146.0); their distributed licenses are retained. The opt-in bedrock feature, which the default build leaves out, raises the minimum Rust version to 1.94.1. Other providers use native HTTP, while the claude-cli backend launches the user's separately installed official CLI. Python SDKs are development references and are not runtime provider dependencies.

Chinese query tokenization uses MIT-licensed jieba-rs 0.11.0, with the HMM parameters from the upstream-compatible jieba-py 0.46.12 release. The source probabilities are data, not Python execution; both licenses are retained under licenses/. The native adapter preserves the reference ASCII boundaries inside mixed Chinese/code terms. Builds without the chinese feature retain Graphify's deterministic bigram fallback.

Audio transcription (the opt-in transcription feature) links whisper-rs 0.16.0 (Unlicense) and whisper.cpp (MIT), using CPU inference, and decodes audio through the user's separately installed ffmpeg executable. Their licenses are retained. Whisper weights are fetched only when transcription needs them, from a pinned public model revision with size and SHA-256 validation. This replaces the Python faster-whisper/CTranslate2 backend; identical transcripts, quantization, speed and model accuracy are not claimed across those inference engines. The optional URL-audio path invokes the separately installed official yt-dlp executable; it is not a Rust rewrite of yt-dlp or a bundled Python interpreter.

URL ingestion uses the MIT-licensed html2text 0.15.5 and its native HTML5 parser in place of optional Python markdownify. The original basic-strip mode remains available. Rich Markdown layout may differ between these converter libraries; source URLs, capture metadata, links and content remain observable in the saved document. The dependency license is retained under licenses/.

Selected post-pin Graphify fixes retain the same upstream notices. Their PR heads, release comparison, local adaptations and regression evidence are recorded in migration/upstream-review/REVIEW.md and adaptations.json; the adapted files, by upstream file, are listed in migration/ADAPTED.json.

Graph adjacency stores single-edge groups inline with smallvec (MIT or Apache-2.0) and uses randomized foldhash (Zlib) for internal lookup tables. The standalone executable uses mimalloc and libmimalloc-sys (MIT) as its Rust allocator; library consumers retain their own allocator choice. These dependency notices are retained under licenses/, with versions in Cargo.lock.

Precise JavaScript/TypeScript resolution (the precise-js feature, in the default build; used with --precise) links the MIT-licensed oxc crates (oxc_parser, oxc_semantic, oxc_ast, oxc_allocator, oxc_span 0.139.0) and oxc_resolver 11.23.0 (MIT). The SCIP importer reads the Apache-2.0 SCIP index format (sourcegraph/scip, scip.proto) with its own decoder; no SCIP code is included. Grammar packs are loaded with libloading (ISC) and built from the same pinned grammar crates as the default build; the pack manifest.json records each crate and version, and each pack is under its grammar crate's licence. Release archives ship no grammar packs.

The licence texts of the Rust crates linked into a release binary are in licenses/THIRD_PARTY_LICENSES.html, generated from Cargo.lock by cargo about (see CONTRIBUTING.md). Model weights, external executables (ffmpeg, yt-dlp, the Claude CLI) and graph database servers are not bundled.