← Volver a las guías

Vision APIs in Production: Image Input Walkthrough

What multimodal image input enables

Multimodal models can interpret text and images in the same request. That makes them useful for invoice extraction, product inspection, screenshot analysis, document understanding, visual search, and support workflows. A production integration needs more than a successful demo: you must select a vision-capable model, transmit images in a supported format, protect user data, validate responses, and control per-request cost.

1. Create an API connection

Start by creating a 59API account, adding credit, and generating an API key. Store the key as an environment variable such as OPENAI_API_KEY; never place it in browser JavaScript, a mobile app, or a public repository.

59API provides pay-as-you-go access to native, official-quality Claude and GPT models without a model downgrade. It is also among the lower-cost relay options, which is useful when image workloads become frequent. Choose a currently listed model that explicitly supports image input. Claude Opus, Sonnet, and Haiku variants may differ in capability and price, so check the model catalog before deploying. If you want to reduce the initial spend, you can sign up for 59API and start with a small test balance; its referral rebate can further reduce ongoing costs.

2. Send an image with an OpenAI-compatible SDK

Install the OpenAI Python package with pip install openai. Configure the SDK to use the 59API host, https://api.59api.com. If your SDK requires a versioned OpenAI-compatible path, use https://api.59api.com/v1 according to the relay documentation.

A basic request contains a text block and an image_url block in the same user message. The structure is: client.chat.completions.create(model="your-vision-model", messages=[{"role":"user","content":[{"type":"text","text":"Read the invoice number and total. Return valid JSON."},{"type":"image_url","image_url":{"url":"https://example.com/invoice.jpg"}}]}]). Replace the model placeholder with the exact vision model name shown in your 59API account.

The image URL must be reachable by the API service over HTTPS. For private or local files, convert the bytes to a base64 data URI such as data:image/jpeg;base64,... instead of exposing a permanent public URL. Set the correct MIME type for JPEG, PNG, WEBP, or another format supported by your selected model.

3. Design prompts for reliable results

Tell the model what to inspect, how to handle uncertainty, and what output your application expects. For extraction, specify field names, data types, and missing-value behavior: “Return invoice_number, invoice_date, currency, subtotal, tax, and total. Use null when a field is unreadable.” Ask the model to distinguish visible evidence from inference rather than guessing.

For multiple images, send several image blocks and label them in the text, such as “Image 1 is the front; Image 2 is the back.” If the model supports image detail controls, use a lower-detail setting for thumbnails and a higher setting for small text. Resize oversized images before upload, but keep enough resolution for the information being read.

4. Move from demo to production

• Validate uploads: Check MIME type, file size, dimensions, and decompression limits on your server. Reject unexpected formats and scan files before processing.
• Protect privacy: Remove unnecessary EXIF metadata, use short-lived storage URLs, define retention rules, and avoid logging raw images or sensitive prompts.
• Keep keys server-side: Your frontend should upload to your backend or controlled object storage. The backend should call 59API and return only the required result.
• Validate output: Treat model responses as untrusted text. Parse JSON, check required fields and types, and route low-confidence or malformed results to human review.
• Handle failures: Set timeouts, retry only transient errors with exponential backoff, and record the model, latency, status, and usage metadata for debugging.
• Control spend: Enforce per-user limits, resize images, cache repeated analyses when appropriate, and select a less expensive capable model for routine jobs.

5. Test before launch

Build a test set containing clear images, blurry photos, rotated documents, empty scenes, handwritten text, and adversarial instructions embedded inside images. Measure extraction accuracy, refusal behavior, latency, and cost—not just whether the request returns HTTP 200. Compare at least two vision-capable models on the same set, then deploy behind a feature flag with monitoring and a human fallback.

With the OpenAI-compatible interface, your application can keep the same client architecture while switching between supported Claude and GPT models through 59API. That makes it practical to start cheaply, learn from real traffic, and scale the image workflow without rewriting the integration.

Precios · Docs