VLM Run

@VLM Run extract structured data from this invoice PDF using the 'invoice' domain schema and return line items with totals

Extraction run started (ID: run_abc123). The invoice domain schema will parse vendor details, line items, subtotals, tax, and total amount. Polling the run shows completion in ~8 seconds. Extracted JSON includes 7 line items with descriptions, quantities, unit prices, and a final total of $3,847.52. All currency fields are normalized to USD with two decimal precision.

This example shows domain-based extraction—VLM Run ships with pre-built schemas for invoices, receipts, forms, and other common documents. Asynchronous runs consume account credits; you must poll the run ID until status is 'succeeded'. Large multi-page PDFs take longer than single images.

@VLM Run create a skill from this zip file that teaches the model how to read our custom lab report format, then use it to extract patient results

Skill 'lab_report_v1' created from the uploaded zip (skill_xyz789). The skill now appears in your account's skill library. Applying this skill to the patient report image, extraction completes in 5 seconds and returns structured fields: patient ID, test date, biomarker panel (12 values with reference ranges), and technician signature. The custom skill captures domain-specific terminology your team defined in the training examples.

Skills let you teach VLM Run your organization's document formats without writing code. You upload examples (zip, prompt, or chat session), and the platform fine-tunes extraction behavior. Skills are reusable across runs but require upfront setup. This is VLM Run's differentiator for non-standard documents.

@VLM Run execute the 'competitive_analysis' agent on these three competitor landing pages and summarize their pricing tiers in a comparison table

Agent execution started (ID: exec_def456). The 'competitive_analysis' agent processes the three screenshots, extracting pricing tables, feature lists, and CTAs. After 12 seconds, the run returns a markdown table with five pricing tiers per competitor, highlighting differences in storage limits, user seats, and API access. The agent also flags that Competitor B hides enterprise pricing behind a contact form.

Agents are multi-step workflows that chain VLM reasoning with tool calls. This example shows agentic synthesis over multiple inputs—useful when you need cross-document comparison or multi-modal reasoning. Agent runs are asynchronous and credit-metered; complex workflows (video, many images) can take minutes. Check account quotas before large batch jobs.

When VLM Run beats manual data entry for finance teams

A 6-person finance team closing books on 200+ vendor invoices every month hits the wall on manual Quickbooks entry. VLM Run's Extract Structured JSON tool pulls line items, totals, and tax fields from PDF or image invoices into a schema you define once, then reuse. The domain discovery tool shows pre-built schemas for common invoice formats; if yours is non-standard, you upload five examples and create a skill that locks in your extraction logic. Async runs mean you queue the batch at 4pm and review structured JSON by morning. This works when invoice volume justifies the API-key setup and per-document credit cost—if you're under 50 invoices a month, a VA is still cheaper. For teams drowning in paper at scale, VLM Run turns a two-day close into a half-day reconciliation check.

Using VLM Run agents to route visual bug reports

A 3-person support team at a SaaS startup gets 40 tickets a day, half with screenshots of broken UI states. The Execute Agent tool runs a VLM agent over each image to classify severity (cosmetic vs. blocker), extract visible error codes, and suggest the right engineering squad. You configure the agent once with a prompt template and route logic; the agent reads the screenshot, checks for known error patterns using a skill you trained on past tickets, and writes a structured triage note. Polling the run ID gives you results in under 10 seconds for most images. This makes sense when visual context is the primary signal and your team wastes more than an hour a day deciding who owns what. If tickets are mostly text or your volume is under 20/day, a simpler keyword router beats the API overhead.

When VLM Run's multimodal extraction handles mixed-format archives

A 5-person compliance team audits 80 vendor contracts every quarter—some are scanned PDFs, some are Word exports, a few are video walkthroughs of SaaS admin panels. VLM Run's Extract Structured JSON tool handles images, documents, video, and audio in one API, so you don't juggle three different OCR vendors. You define a schema for liability caps, renewal dates, and data-residency clauses, then batch-upload the archive. Video and document runs are async; you poll with Get Run until all extractions finish, then diff the structured output against your compliance checklist in a spreadsheet. This wins when format chaos is the blocker and your team has API-key access to provision the integration. If all your contracts are already in a single text format, a cheaper document-parser MCP does the job without the multimodal tax.