The semantic index was ~90 896 per-entry files (index/embeddings/<safe-model>/
<sha256>.f32), a deployment liability: the file count — not the 356 MB — is
hostile to container layers, object storage, and Git-LFS, and Pipeline::load
opened every one individually. This packs a complete per-file store into one
self-describing matrix file loaded in a single read.
New src/packed.rs owns the format: 4-byte magic "AIPK", u32 LE format_version,
u32 LE header length, a JSON PackedHeader {n_rows, dim, embed_model,
corpus_sha256, corpus_name}, pad to a 4-byte boundary, then row-major
little-endian f32 payload in corpus order (row i = corpus entry i). write_packed
is atomic (temp+rename); read_packed rejects a wrong magic, a newer-than-
supported version, a malformed header, or a payload length inconsistent with
n_rows×dim. corpus_sha256 hashes the raw corpus file bytes.
Pipeline::load now calls a free fn load_packed_or_store(cfg, entries) that
prefers the packed file and falls back to today's all-or-nothing per-file store,
reporting which path it took via a new IndexStatus {Ok, Absent, Mismatch}
carried into Diagnostics. `alpha-id index --pack` writes the file from a
complete store (refusing an incomplete one with exit 2), recording corpus_sha256
+ embed_model in the header as the artifact-binding data.
Iteration boundary (deliberate): this is iteration 1 of the packed-versioned-
bundle cycle. load_packed_or_store validates the packed file STRUCTURALLY only
(magic/version/payload length via read_packed, plus n_rows == entries.len());
it does NOT yet compare the header's corpus_sha256/embed_model against the live
config, and IndexStatus::Mismatch fires only on structural corruption. The
load-time mismatch warning, the index= token on the diagnostics line, and the
Hybrid-degrade-names-reason wiring are iteration 2 (issue #4), which consumes
the binding data this commit writes. The per-file store is untouched — it stays
the build/resume path; the packed file is an additive serve-path artifact.
mmap was considered and rejected: the brute-force full-scan cosine search
touches every row per query, so memory-mapping yields no runtime benefit over a
single read and would add an unsafe dependency (spec § Out of scope).
Verification: cargo test green (61 tests; 9 new — 3 format round-trip/hash/bad-
magic, 4 load-path incl. packed-beats-store and structural n_rows mismatch, 2
--pack CLI incl. exit-2 refusal). An end-to-end seam test writes the packed
file via the CLI, removes the per-file store, and asserts the pipeline builds
its index from the packed file alone as IndexStatus::Ok (#3 acceptance b).
Spec: docs/specs/2026-05-31-packed-versioned-index-bundle.md
Plan: docs/plans/2026-06-01-packed-versioned-bundle-iter1.md
closes#3
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Bite-sized TDD plan for iteration 1 of the packed-versioned-bundle cycle:
the on-disk packed format (src/packed.rs), the IndexStatus enum + packed-first
Pipeline::load path with per-file-store fallback, and the `index --pack`
subcommand. Structural validation only (magic/version/n_rows/payload length);
the semantic corpus+model mismatch enforcement, load warning, and index=
diagnostics token are deferred to iteration 2 (#4), as documented in the
plan's iteration-boundary note.
refs #3
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Design spec for the "versioned bundle" cycle binding issues #3 (pack the
per-file embedding store into one contiguous matrix file) and #4 (bind
{corpus snapshot, embed model, index} as one versioned artifact).
Ratified decisions (user-approved across the brainstorm):
- Approach 1: single self-describing file (magic + JSON-header manifest +
row-major f32 payload), loaded with one fs::read. No mmap (a brute-force
full-scan cosine index touches every row per query, so mmap yields no
runtime benefit and is rejected to avoid an unsafe dependency). VectorIndex
and the per-file store are untouched — the store stays the build/resume
path, the packed file is an additive serve-path artifact.
- Mismatch posture: loud but not fatal. A structurally wrong index (corpus
sha256 or embed_model disagreement) builds no VectorIndex and surfaces as a
distinct IndexStatus::Mismatch in diagnostics plus a load-time warning,
instead of silently degrading Mode Hybrid to Lexical. Mode Lexical, which
never needs the index, keeps serving.
- Two iterations, one format: iter 1 delivers the format + `index --pack` +
packed-first load; iter 2 enforces the manifest check and the observable
index_status surface.
grounding-check: PASS (7 load-bearing assumptions ratified by green tests).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The retrieval-mode enum used opaque single letters `Mode { C, A }`;
neither the identifier nor the operator-facing `--mode a` token
conveyed what the mode does. Rename to self-describing names that map
directly onto the glossary definitions:
Mode::C -> Mode::Lexical (BM25 + tag filter, offline, no IONOS)
Mode::A -> Mode::Hybrid (lexical u semantic -> RRF -> rerank)
Decisions taken (the issue's open questions, resolved):
- CLI canonical tokens are now `lexical` / `hybrid`; the default is
`lexical`. The legacy `c` / `a` tokens stay accepted as
case-insensitive input aliases (zero cost; protects hand-typed
invocations). A unit test pins the alias contract.
- The diagnostics `mode` string (Debug-derived) now emits
"Lexical"/"Hybrid". Verified safe: no downstream consumer parses it.
doctate is the future integrator and embeds the library (the `Mode`
enum), not the JSON string (design spec §103).
- The reserved, out-of-scope generative tier (spec's "Mode B") keeps
conceptual room: a future `Mode::Generative` does not collide with
the retrieval-strategy names Lexical/Hybrid.
Surface: enum + FromStr, CLI defaults and `eval --mode both`
expansion, the private `collect_mode_c`/`collect_mode_a` methods (now
`collect_lexical`/`collect_hybrid`) and their comments, the two
pipeline mode test files (renamed), the eval CLI test, the glossary
(old letter forms demoted to Avoid lists), and the spec CLI synopsis.
closes#2
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Records the ratified naming decisions (Lexical/Hybrid, c/a kept as
input aliases, diagnostics string free to change — no downstream
consumer) and the bite-sized rename task list for the implement step.
refs #2
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The --chapters filter compared the raw CLI token by exact string
equality against the zero-padded chapter tags ("01".."22") that
the corpus carries (roman_to_padded, src/claml.rs). The CLI only
split and trimmed the input, so the intuitive forms — `--chapters 4`
or `--chapters E,I` — never matched any candidate and the suggestion
list came back empty, indistinguishable from "no clinical match".
Add a pure `normalize_chapters` helper that zero-pads bare 1-2 digit
chapter numbers to two digits and partitions the token list into
known chapters (1..=22) and unrecognised tokens. The Suggest arm now
feeds the padded set into the filter and emits a stderr warning for
each unrecognised token, so misuse is visible instead of silent.
`--chapters 4` and `--chapters 04` now behave identically. ICD
code-letter mapping (E -> chapter) is deliberately left out of scope:
under this contract a letter is simply an unrecognised token and
triggers the warning.
closes#1
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pins the property that the `--chapters` CLI token list normalizes
to the zero-padded chapter-tag form ("01".."22") that Filter::accepts
compares against, so "4" filters identically to "04", and any token
mapping to no known chapter is surfaced as unrecognised (for a stderr
warning) rather than silently collapsing the result set to empty.
RED: alpha_id::model::normalize_chapters does not exist yet — the
test fails to compile. The GREEN side follows in the next commit.
refs #1
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add .claude/dev-cycle-profile.yml (skills-plugin profile) with
paths.glossary pointing at the new docs/glossary.md, so the glossary
becomes standing reading for every role. The glossary pins 19 canonical
terms bootstrapped from the design spec and source doc comments;
canonical forms are English (repo-English rule), German spec-prose forms
listed under Avoid.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The superpowers workflow methodology is no longer in use. Move the
methodology-neutral design spec to docs/specs/ as the project's design
ledger, drop the superpowers-bound implementation plan (it referenced
superpowers sub-skills), and remove the now-dead /docs/superpowers/
.gitignore entry.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Emit a `… embedded N/total (pct%)` line every 1000 items (and at
completion) so a full --confirm run is observable rather than silent
for its full duration.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
split_segments now treats each non-empty trimmed line as one retrieval
unit; blank lines are skipped, not paragraph delimiters. Name/signature
retained (minimal-churn decision). Full suite 45/0. End-to-end: feeding
the GP dictation as extracted phrases recovers I10.90 (rank 4) which was
entirely missing in both modes under the paragraph contract, and largely
dissolves the cross-segment diabetes crowding.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Source-string analysis showed the Hypertonie miss is not a bug but a
structural consequence of embedding whole noisy Anamnese paragraphs
against terse Alpha-ID corpus strings. Resolution (user-decided): the
caller (doctate's LLM) extracts diagnostic phrases; this project still
contains no LLM — it just receives a cleaner input contract.
Spec: new §1a (rationale + generous extraction contract + the honest
extractor-is-recall-bottleneck trade-off); §1/§2/§3/§6/§8/§9/status
updated; §6 reworks eval (representativeness shift, paraphrase axis,
hand-labeled phrase-list set, separated extraction-ceiling vs matcher
recall). Plan: Goal/Architecture, revision banner, Task 6 rewritten
(one non-empty line = one unit; split_segments name/fields retained,
minimal-churn decision), Task 11/19 revision notes.
Code not yet changed (segment.rs still blank-line splits) — next step.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Smoke-testing invented German dictations exposed that `ae`/`oe`/`ue`
spellings collapsed lexical recall (e.g. "Hyperlipidaemie" → garbage,
"Hyperlipidämie" → correct). normalize() now folds ä→ae, ö→oe, ü→ue,
ß→ss; it is applied only in lexical.rs (both corpus index and query),
so the path is symmetric under either spelling. The embedding path
(raw text for bge-m3) is intentionally untouched.
eval.rs inject_errors case 2 changed from umlaut-deletion (ä→a) to the
realistic ASCII-digraph variant (ä→ae) so the eval harness exercises —
and guards against regression of — exactly this failure class.
TDD: failing tests first; normalize_tests 5/5, full suite 44/0.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Documents the corpus↔ClaML 5th-digit gap, the decided Modifier/
ModifierClass expansion fix (Goal 1), and the deterministic /
seeded-sample eval corrections from the final review.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Fix 1: capture <ValidModifierClass> codes from <ModifiedBy all="false">
and filter each modifier level to only valid codes before cartesian
product; eliminates ~194 phantom codes (e.g. M07.01–03) while keeping
valid ones (M07.00/04/07/09). Split Start/Empty XML event arms to
correctly track children of <ModifiedBy>.
- Fix 2: synthesized para295/para301 inherit parent values; "V" is
promoted to "P" (parent "V" = needs subdivision, terminal is codeable).
Other types ("P", "O", "Z", "") pass through unchanged.
- Fix 3: debug_assert in ModifierClass handler guards against ClaML
document-order assumption changing silently in debug/test builds.
- Fix 4: add expands_single_modifier_chain regression test verifying
C88.00/C88.01 synthesis from a single all="true" ModifiedBy (corpus-
verified: Alpha-ID I30531, I31044).
ICD-10-GM models the 4th/5th sub-digits of many groups (incl. all E10-E14
diabetes) via <Modifier>/<ModifierClass> referenced by ordered <ModifiedBy>
on the parent <Class>, not as nested <Class> elements. load() previously
returned nothing for codes like E11.90/E11.72/I10.91, so ~1.7k corpus
codes (incl. ~1.2k billable) fell back to the non-billable 3-digit root
and were silently dropped under --billable-only.
Collect per-Class ModifiedBy refs and all ModifierClass defs during the
existing single stream, then expand: cartesian product of the referenced
modifiers' ModifierClass codes in ModifiedBy order, concatenated verbatim
onto the parent code (E11 + .9 + 0 = E11.90), matching the Alpha-ID
corpus format exactly. Terminal leaves are billable (Para295/Para301=P);
description = parent label + each ModifierClass label joined by ' - ';
chapter/group/exotic/ifsg/content inherited from parent. Purely additive:
explicit <Class> entries always win. Generic over every modifier group.
Map size 12334 -> 17249 (+4915 synthesized).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Implement Pipeline::load/suggest for Mode C: lexical search per segment,
cross-segment dedup via fusion, tag derivation, filter application.
Add meta_lookup fallback (exact → strip trailing 0 → 3-char root) to
bridge the ICD code granularity gap between the corpus (E11.90-style
6-char codes) and ClaML (which only stores E11, E11.0-style entries).
Brainstormed design: free-text dictation (marker-segmented by
doctate) -> single prioritized top-10 ICD-10-GM list with tag
filtering. Two modes (C lexical baseline, A hybrid retrieval +
IONOS bge-m3 + Qwen3-VL-Reranker-8B), eval harness via synthetic
transcription errors, one-time cost-guarded embedding.
Licensed BfArM data dirs and local settings are gitignored.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Brummel