The header referenced the license identifier of a third-party library's
bundled dependency while explaining that the implementation deliberately
does not use that library. License scanners match the bare identifier
string and misclassify the file. Reword to state provenance directly:
written from the published papers, nothing vendored or derived.
Comment-only change; also picks up the formatter pass on this file.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
The semantic pass kept four dense 768-float vectors resident per function
(ri/api/type/deco = 12.3 KB per func, ~9.4 GB transient on the linux kernel)
solely so candidate pairs could be cosine-scored after LSH. Replace the
resident floats with 4-bit quantized codes (~0.5 KB each, ~6x): new
src/semantic/rotsq.{h,c} — rotated scalar quantization from the RaBitQ
family of algorithms (Gao et al., SIGMOD 2024/2025), implemented from the
papers in plain C11 (no Eigen/library dependency): a deterministic
XXH3-seeded +-1 diagonal followed by a Fast Walsh-Hadamard rotation
(768 -> 1024 padded) spreads vector mass so per-vector scalar quantization
is near-optimal; the inner product is recovered from the codes with the
exact SQ expansion (one u8 integer dot + four multiplies) — a deterministic
pure function of the codes. Named rotsq, not rabitq: this is the family
core, not the papers' exact codebook construction.
Producers build each dense vector in a worker-local buffer, encode, and
discard; the int8 qvec export is unchanged. LSH signatures are computed
from DEQUANTIZED codes against hyperplanes in the rotated basis — random
hyperplanes are basis-agnostic, so no dense originals are retained and the
phase structure is untouched.
Estimator quality gated by sem_rotsq_ip_error_bounds (seeded vectors:
mean error < 1%, max < 4% of unit scale, ASan-green). Effect on output:
multi-threaded determinism preserved (xfs: two runs byte-identical);
SEMANTICALLY_RELATED edges shift once as scores move within ~1% of the
0.75 threshold (xfs: ~130 -> 203; sampled new pairs are high quality —
alloc/free, read-verify and collapse/insert families).
Suites: semantic 33 (incl. the new estimator gate), simhash 24,
pipeline 216 — zero failures.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
Per-function complexity metadata is now stored on graph nodes and queryable,
alongside several indexing performance and correctness fixes developed and
validated together (3704 tests, ASan/UBSan clean).
Bottleneck metrics (query via query_graph):
- Tier A (in the extraction AST walk): cyclomatic (complexity), cognitive
(nesting-weighted), loop_count, loop_depth (max nested-loop depth),
param_count, max_access_depth.
- Tier B (new pre-dump pass, pass_complexity.c): transitive_loop_depth
propagated along CALLS edges + a recursive flag (direct self-recursion and
mutual-recursion cycles), plus the call-context signals linear_scan_in_loop,
alloc_in_loop, recursion_in_loop and unguarded_recursion.
- query_graph and get_architecture tool descriptions document the metrics and
the Leiden community clusters.
Cypher engine:
- node_prop exposes arbitrary persisted node properties to WHERE/RETURN.
- Fix projection aliasing: multi-property rows shared a single static buffer so
every column returned the last value read; now per-column/rotating buffers.
- Fix a stack-use-after-scope in aggregate RETURN (caller-owned value buffers).
Indexing performance:
- Gate C/C++ #define Macro-node extraction to full mode (it is ~49% of nodes on
the Linux kernel); moderate/fast skip it.
- Emit the complexity property block only for Function/Method nodes so the
millions of Macro/Field/Variable/Class/Enum nodes no longer carry zeroed
fields — large RAM reduction at scale.
- Classify node types via tree-sitter TSSymbol bitsets in cbm_kind_in_set
instead of per-node strcmp scans (thread-local cache, strcmp fallback;
behaviour-identical).
- Subsample frequent (Zipfian) tokens in the semantic co-occurrence finalize;
~14x faster finalize on the kernel, output unchanged.
- pass_lsp_cross: replace O(n^2) linear dedup with hash-set dedup.
Windows:
- Canonicalize drive-letter case during path normalization so "c:/repo" and
"C:/repo" derive the same project key and cache file (#394/#227/#367).
Tests: extraction, pipeline and cypher regressions covering all of the above.
- semantic.c: free strdup'd keys+values in token_map via cbm_ht_foreach
before cbm_ht_free (44KB leak in 7920 allocations)
- code_vectors_blob.S: add Windows COFF branch (.rdata,"dr") alongside
macOS Mach-O and Linux ELF sections
- Parallelize pass_similarity and pass_semantic_edges via worker pool with
thread-local edge buffers; sequential final merge since gbuf is not
thread-safe. Adds cbm_lsh_query_into() as a thread-safe variant with
caller-provided candidate buffer.
- Add activatable profiling subsystem (CBM_PROFILE=1 env or --profile flag)
for step-level timing of extract, resolve, corpus build, vector phases,
and sqlite dump. Zero overhead when disabled.
- Fix cbm_index_mode_t enum mismatch between pipeline.h (FULL=0, MODERATE=1,
FAST=2) and discover.h (FULL=0, FAST=1). mode=fast silently no-op'd
fast-discovery filtering because discover.c compared against the wrong
value. Linux kernel fast mode went 1:40 -> 3:11 as a result; now back to
1:40. Broaden the filter guard to mode != CBM_MODE_FULL so MODERATE and
FAST both get aggressive discovery.
- Clamp cbm_sem_combined_score output to [0, 1]. The proximity multiplier
returns up to 1.10 as a same-file boost which could push the final
cosine score above 1.0.
- Short-circuit semantic scoring when MinHash jaccard >= 0.95. Exact
near-clones are already emitted as SIMILAR_TO edges; returning 0 here
avoids flooding SEMANTICALLY_RELATED with cross-service copy-paste
boilerplate and frees the edge budget for genuine vocabulary-bridged
relations.
- Validate search_graph semantic_query as an array of strings and return
a clear error for a single-string input. Update the tool description
to spell out the requirement explicitly with an example.
- JSON-escape user-controlled strings (callee names, call arguments,
URL paths, import local_name) in call/argument properties. Introduces
cbm_json_escape() in foundation/str_util.
- Skip SQLite pending_byte_page (file offset 0x40000000) during raw page
writes in sqlite_writer to avoid corrupting databases that cross the
1 GiB boundary.
- Migrate pretrained vector blob from UniXcoder (51K tokens) to
nomic-embed-code (40856 tokens x 768d int8). Includes the extraction
script under scripts/extract_nomic_vectors.py.
Major quality improvement: replace random index vectors with pre-trained
Nomic nomic-embed-code token embeddings. Tokens like 'error' and 'exception' now start
with similar vectors (learned from millions of code repos) instead of
arbitrary random projections. Co-occurrence enrichment adds project-specific
context on top.
Architecture:
- vendored/nomic/code_vectors.bin: 37.7MB raw int8 vectors
- vendored/nomic/code_vectors_blob.S: assembler .incbin (instant build)
- vendored/nomic/code_vectors.h: extern declarations + pretrained_vec_at()
- vendored/nomic/code_tokens.h: 40856 token strings (575KB)
- semantic.c: cbm_sem_random_index() now looks up pretrained vectors first,
falls back to sparse random for unknown tokens
- CBM_SEM_DIM raised from 256 to 768 to match Nomic nomic-embed-code
Also: RRI (Reflective Random Indexing), code pattern vocabulary injection,
120+ abbreviation expansions, callee/caller/body token enrichment,
label filter (Function/Method/Class only) in vector search SQL.
Binary size: 136MB → 169MB (+33MB from embedded vectors).
Search quality: keyword queries return relevant error-handling functions.
Domain-specific keyword queries return the expected functions.
Martin Vogel