pub const DEFAULT_TRANSCRIPT_MODEL: &str = "gemini-3.5-flash-lite";Expand description
The default model for extracting observations from a transcript.
Smaller than DEFAULT_EXTRACTION_MODEL, because the two jobs are not
equally hard. Reading an utterance that is already in front of you is easier
than canonicalising a question into the English search terms the stored
fact was canonicalised into — planning has only the question to go on, and
that is where a smaller model actually degrades.
Measured by holding observations at gemini-3.5-flash-lite and varying only
the plan model, over code_switched_e2e’s cross-lingual retrieval case
(a Hinglish question against an English-canonicalised fact), n=10 runs each:
| Plan model | Passes |
|---|---|
gemini-2.5-flash | 8/10 |
gemini-3.5-flash-lite | 3/10 |
The fact stores correctly either way — it is the question that fails to canonicalise, so the query and the record never meet. Ingestion showed no such gap, which is what makes the split worth having rather than just downgrading everything.
Latency, from model_latency_probe (p50):
gemini-2.5-flash | gemini-3.5-flash-lite | |
|---|---|---|
| observation extraction | 2144 ms | 1115 ms |
| prepare incl. model plan | 1812 ms | 1150 ms |
Note the 2/10 residual: this case is flaky under every configuration
including the previous all-gemini-2.5-flash default. Treat these as rates,
not verdicts, and do not read a single green run as a fix.