Repository navigation
Extract BaseAPIDescriptionScorer shared by LLMDescriptionScorer and TypeSafeDescriptionScorer #357
Description
After #350, LLMDescriptionScorer (llm_encoder.py, 319 lines) and TypeSafeDescriptionScorer (typesafe.py, 416 lines) carry the same machinery side by side:
| concern | LLMDescriptionScorer |
TypeSafeDescriptionScorer |
|---|---|---|
max_concurrent / max_per_second params, sync client when max_concurrent is None |
:111-136 |
:160-209 |
_init_event_loop + run_until_complete(aiometer.run_all(...)) fan-out |
:247-252, :282-291 |
:331-340, :259-270 |
| per-utterance sync/async ask with failure → uniform row | :202-224, :264-268 |
:342-362, :303-307 |
clear_cache dropping clients + loop |
:274-280 |
:384-388 |
dump / load with exclude=[asyncio.BaseEventLoop, dict] and a JSON sidecar for the config |
:293-318 |
:390-416 |
They differ only in what one request is (an OpenAI structured-output prompt vs a TypeSafe Choice/Noul question), how its answer maps to a similarity row, and the cache key.
Proposed
BaseAPIDescriptionScorer(DescriptionScorer) owning the shared part: the request fan-out (_ask_many), the event-loop helper, cache-first lookup, the fatal/transient error policy, clear_cache, dump/load. Subclasses implement _ask_one_async(utterance) -> row, _cache_key(utterance), client construction, and the similarity mapping. description_bi / description_cross stay on DescriptionScorer as they are.
Deliberately left out of #350 to keep it reviewable. Best done after the three behavioural fixes so each policy is written once, in the base: #351 (error policy), #353 (event loop), #354 (cache before limiter).
Activity