Three memory quick-wins from the gbuf compaction research (graph output
verified identical; all delete/incremental paths suite-covered):
1. Release ALL lookup indexes right after build_dump_nodes instead of after
edge building: nothing later in the dump reads them, and at kernel scale
they are ~3.8 GB (hash buckets + key strings + pointer arrays) that
previously coexisted with every dump-side transient — the dump-phase RSS
peak. At 2x-kernel scale this is ~7.5 GB at exactly the OOM point.
2. node_by_id was a hash table keyed on STRDUP'D DECIMAL STRINGS of the
int64 id — ~0.44 GB of buckets and key strings, plus a snprintf + strdup +
hash on every hot find_by_id call. Ids are dense and sequential; a plain
grow-on-demand pointer array replaces all of it with one bounds-checked
load.
3. The per-node and per-edge `project` field was write-only (every reader
uses the store structs or gb->project directly; -Werror confirms) —
8 bytes x 24M instances ~= 192 MB of dead pointers, deleted.
Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>
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.
Vector search pipeline (end-to-end):
- store.c: cbm_cosine_i8 custom SQLite function (int8 dot product with
sqrt magnitude), registered at store open time
- store.c: cbm_store_vector_search() builds merged query vector from
keywords via base RI vectors (xxHash), int8-quantizes, runs SQL query
with JOIN nodes ON node_id = id for metadata
- mcp.c: search_graph gains semantic_query param (array of strings),
returns semantic_results alongside regular results
- sqlite_writer.c: fixed build_vector_record to include node_id in
record body (matches build_node_record pattern)
- mcp.c: use yyjson_mut_obj_add_strcpy (not add_str) to avoid
use-after-free when vector results are freed before JSON serialization
Quality note: base RI vectors at query time don't capture co-occurrence
enrichment — results are moderate quality. Storing enriched token vectors
would improve this significantly (future work).
Incremental pipeline (pipeline_incremental.c):
- Disk-based delta re-indexing: classifies files by mtime+size vs stored
hashes, deletes changed nodes (edges cascade), re-parses only changed
files, merges into existing DB via cbm_gbuf_merge_into_store
- pipeline.c auto-routes: existing DB with hashes -> incremental, else full
- Full pipeline persists file hashes after dump for next incremental run
- No-op when nothing changed (watcher polls every 5-60s, 99% skip)
- 4 tests: noop, changed file, deleted file, new file added
ADR hints (ported from Go):
- index_repository: adr_present + adr_hint when no ADR exists
- get_graph_schema: adr_present + adr_hint per project
- manage_adr GET: adr_hint with creation instructions when no ADR
Complete C implementation of the indexing pipeline (src/), parallel worker
pool, graph buffer with merge support, SQLite writer, and 1893-test suite.
Linter setup: zero warnings from clang-tidy (all checks enabled, no NOLINT
suppressions), cppcheck, and clang-format. All issues fixed at source —
proper headers for external linkage, GROW_ARRAY macro restructured to
eliminate type parameter, null-deref paths guarded, named intermediates
for suspicious-argument checks.
Martin Vogel