cookbook/data_labeling/_03_text_extraction/TEST_LOG.md
Tested 2026-07-18 against gemini-3.5-flash, agno 2.7.4.
Status: PASS
Description: Extracts a flat Contact (name, email, phone, company, title) from two email-signature-style samples using output_schema.
Result: Sample 1 extracted all fields verbatim: name='Sarah Johnson', email='[email protected]', phone='+1-555-0102', company='Acme Corp.', title='VP of Marketing'. Sample 2 extracted name='Mike', email='[email protected]' with phone, company, and title left None as instructed.
Status: PASS
Description: Extracts a Meeting containing a list of nested ActionItem objects (owner, description, due_date) from a four-line meeting transcript; vague group asks are to be ignored.
Result: Three action items extracted with correct owners: Mike ('Send out the updated roadmap'), Sarah ('Set up the kickoff with the design team'), Mike ('Draft the budget memo'). Jess's vague 'budget approval at some point' was correctly excluded. All due_date fields were None this run - the transcript only contains relative dates ('by Friday', 'end of next week'), which the model did not resolve to ISO dates.
Status: PASS
Description: Same contact-extraction task with each field wrapped in a ConfidentField (value plus Literal high/medium/low confidence) to support routing low-confidence fields to review.
Result: Sample 1 (full signature) returned all five fields with confidence='high' and verbatim values. Sample 2 ('ping @mike on the eng team') returned name=('mike', high), title=('eng team', medium), and email/phone/company as (None, low). Confidence spread is sensible, though name='mike' at 'high' and 'eng team' as a title are looser calls this run.