Gemini Vision Extract

Use this skill when you have an image and need structured fields out of it - line items off a receipt, totals off an invoice, values off a chart - rather than a paragraph of description. The Gemini call is constrained with response_mime_type=application/json plus a Pydantic response_schema, so the output parses into a typed object every time.

When to invoke

• User says: "read this receipt", "parse the invoice image", "OCR into structured data", "extract fields from a screenshot".
• Code in the conversation feeds an image to an LLM and then regex-scrapes prose for values.

When NOT to invoke

• The source is already machine-readable text/HTML/PDF-with-text - extract directly, no vision model needed.
• You need pixel-perfect OCR of dense documents at scale; a dedicated document-AI service may fit better.

Concrete example

User input:

Pull merchant, date, and total out of these receipt photos into JSON.

Output:

from vision_extract import Receipt, extract   # GEMINI_API_KEY in the environment

data = extract("receipt.png", schema=Receipt)
print(data.model_dump_json(indent=2))
# {"merchant": "Corner Cafe", "date": "2026-06-06", "total": 4.5, "currency": "USD", "items": [...]}

Pattern to apply

1. Define a Pydantic model with optional fields and pass it as response_schema; make absent fields nullable.
2. Set response_mime_type=application/json so the model returns parseable JSON, not commentary.
3. Read GEMINI_API_KEY from the environment; never inline it.
4. Prefer resp.parsed (already a validated instance); fall back to validating resp.text if the SDK returns none.
5. Pick the right model tier - Flash for cheap high-volume extraction, Pro when accuracy matters (see [[gemini-flash-budget]]).

Reference: assets/vision_extract.py.

Source

Distilled from production use across the author's automation projects. v1.0.0. See also: [[gemini-structured-output]], [[gemini-flash-budget]], [[gemini-cost-tracker]].

→ Build the full runnable bot with Trawlkit.