personal memory agent
Disable thinking for Gemini transcription to fix max_tokens truncation
Gemini 3 Flash applies a default thinking budget when none is specified,
consuming output tokens from the 16384 limit. Dense 5-minute audio segments
were hitting max_tokens with only ~2000 chars of transcript generated.
Setting thinking_budget=0 gives the full token budget to the JSON output.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+3
+3
observe/transcribe/gemini.py
+3
observe/transcribe/gemini.py
···
321
321
]
322
322
323
323
# Call Gemini via think.models.generate()
324
+
# thinking_budget=0 disables thinking — transcription is extraction, not
325
+
# reasoning, and Gemini's default thinking budget consumes output tokens.
324
326
response_text = generate(
325
327
contents=contents,
326
328
context="observe.transcribe.gemini",
327
329
temperature=0.3,
328
330
max_output_tokens=16384,
329
331
json_output=True,
332
+
thinking_budget=0,
330
333
)
331
334
332
335
transcribe_time = time.perf_counter() - t0