mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-29 19:01:39 +00:00
pabloglzgandClaude Fable 5
164f310374
dimsProviderOptions() threads input_type ('query' | 'document') into
providerOptions.openaiCompatible for asymmetric models (ZE zembed-1,
Voyage v3+), but the AI SDK's openai-compatible adapter validates
providerOptions against a fixed schema and silently drops the field
before building the HTTP body. Every embedQuery() was therefore encoded
document-side: the ZE shim's hard default fired ('document'), Voyage and
local openai-compat servers got no input_type at all, and asymmetric
retrieval silently collapsed toward surface-token overlap — while the
providerOptions-level contract test stayed green.
Fix: an AsyncLocalStorage (same pattern as __budgetStore) populated in
embedSubBatch() only when providerOptions actually threads an
input_type, read at body-rewrite time by the fetch shims:
- zeroEntropyCompatFetch: recovers the threaded value; document default
preserved for ingest paths.
- voyageCompatFetch: opt-in like the dims.ts Voyage branch — inject only
when threaded; the field stays off the wire otherwise.
- NEW openAICompatAsymmetricFetch: fallthrough default for every other
openai-compatible recipe (llama-server, litellm, ollama, ...) — the
canonical local/proxy paths for asymmetric models. Strict pass-through
when nothing was threaded, so symmetric deployments see zero wire
change; recipes with their own compat fetch (azure) keep it via the
compat.fetch ?? precedence.
KNOBS_HASH_VERSION bumped 10→11: cached query_cache rows were keyed on
document-side query vectors; pre-fix rows must not be served to post-fix
lookups (same convention as the v=3 embedding-provider bump). One-time
global cold-miss on upgrade; refills within cache.ttl_seconds.
Tests: test/embed-input-type-wire.test.ts runs the REAL SDK transport
with a mocked global fetch and asserts on the outbound body — the only
layer where this regression is observable. Covers ZE hosted, llama-server,
litellm, ollama (query + document sides) and pins the pass-through for
non-asymmetric models and Voyage's opt-in shape. 4 of the original 7
assertions fail on master, proving the pin. One structural pin in
test/ai/zeroentropy-compat-fetch.test.ts updated to the new line shape
(same semantic); KEY_FILES.md gateway.ts entry updated to the new truth.
Supersedes #1400 (closed unmerged) — same ALS mechanism, extended to
Voyage + all openai-compatible recipes. Credit to @billy-armstrong for
the original diagnosis.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>