ensemble: document-pipeline

agents:
  # Extract text
  - name: extract
    operation: think
    config:
      provider: openai
      model: gpt-4o
      image: ${input.document_url}
      prompt: Extract all text from this document

  # Classify document type
  - name: classify
    operation: think
    config:
      prompt: |
        Classify this document:
        - invoice
        - receipt
        - contract
        - resume
        - other

        Text: ${extract.output}
        Return JSON: { "type": string, "confidence": number }

  # Extract structured data
  - name: extract-invoice
    condition: ${JSON.parse(classify.output).type === 'invoice'}
    operation: think
    config:
      prompt: |
        Extract invoice data:
        ${extract.output}

        Return JSON: {
          "invoice_number": string,
          "date": string,
          "vendor": string,
          "total": number,
          "line_items": []
        }

  # Validate extracted data
  - name: validate
    agent: validator
    inputs:
      data: ${extract-invoice.output}
      schema: ${component.invoice-schema@v1.0.0}

  # Store in database
  - name: store
    condition: ${validate.output.valid}
    operation: storage
    config:
      type: d1
      query: INSERT INTO invoices ...
      params: ${validate.output.data}

ensemble: process-invoice

agents:
  # OCR + extraction
  - name: extract
    operation: think
    config:
      provider: openai
      model: gpt-4o
      image: ${input.invoice_image}
      prompt: |
        Extract invoice information:
        - Invoice number
        - Date
        - Vendor name and address
        - Line items (description, quantity, price)
        - Subtotal, tax, total

        Return as JSON.

  # Validate amounts
  - name: validate-math
    operation: code
    config:
      code: |
        const invoice = JSON.parse(${extract.output});
        const lineTotal = invoice.line_items.reduce((sum, item) =>
          sum + (item.quantity * item.price), 0
        );

        return {
          valid: Math.abs(lineTotal - invoice.subtotal) < 0.01,
          calculated_total: lineTotal,
          invoice_total: invoice.subtotal
        };

  # Check vendor
  - name: check-vendor
    operation: storage
    config:
      type: d1
      query: SELECT * FROM approved_vendors WHERE name = ?
      params: [${JSON.parse(extract.output).vendor}]

  # Require approval for new vendors
  - name: approve-vendor
    condition: ${check-vendor.output.length === 0}
    agent: hitl
    inputs:
      data: ${JSON.parse(extract.output)}
      prompt: "New vendor - approve invoice?"
      approvers: [finance@example.com]

  # Auto-approve known vendors under threshold
  - name: auto-approve
    condition: ${check-vendor.output.length > 0 && JSON.parse(extract.output).total < 5000}
    operation: code
    config:
      code: return { approved: true, auto: true };

  # Store invoice
  - name: store
    condition: ${auto-approve.executed || approve-vendor.output.approved}
    operation: storage
    config:
      type: d1
      query: INSERT INTO invoices ...

ensemble: screen-resume

agents:
  # Extract resume data
  - name: extract
    operation: think
    config:
      provider: openai
      model: gpt-4o
      prompt: |
        Extract from resume:
        ${input.resume_text}

        Return JSON: {
          "name": string,
          "email": string,
          "phone": string,
          "experience_years": number,
          "skills": [string],
          "education": [{ "degree": string, "school": string, "year": number }],
          "work_history": [{ "company": string, "title": string, "duration": string }]
        }

  # Score against job requirements
  - name: score
    operation: think
    config:
      provider: openai
      model: gpt-4o
      prompt: |
        Score resume fit (0-100) for this position:

        Requirements: ${input.job_requirements}

        Resume: ${extract.output}

        Return JSON: {
          "overall_score": number,
          "scores": {
            "experience": number,
            "skills": number,
            "education": number
          },
          "strengths": [string],
          "gaps": [string],
          "recommendation": "interview" | "maybe" | "reject"
        }

  # Shortlist high scores
  - name: shortlist
    condition: ${JSON.parse(score.output).overall_score >= 70}
    operation: storage
    config:
      type: d1
      query: |
        INSERT INTO shortlisted_candidates (name, email, score, data, job_id)
        VALUES (?, ?, ?, ?, ?)
      params:
        - ${JSON.parse(extract.output).name}
        - ${JSON.parse(extract.output).email}
        - ${JSON.parse(score.output).overall_score}
        - ${extract.output}
        - ${input.job_id}

  # Auto-reject low scores
  - name: auto-reject
    condition: ${JSON.parse(score.output).overall_score < 50}
    operation: email
    config:
      to: ${JSON.parse(extract.output).email}
      subject: "Application Update"
      body: |
        Thank you for your interest...

ensemble: analyze-contract

agents:
  # Extract key terms
  - name: extract-terms
    operation: think
    config:
      provider: openai
      model: gpt-4o
      prompt: |
        Extract key contract terms:
        ${input.contract_text}

        Return JSON: {
          "parties": [string],
          "effective_date": string,
          "termination_date": string,
          "payment_terms": string,
          "termination_clause": string,
          "liability_cap": string
        }

  # Identify risks
  - name: identify-risks
    operation: think
    config:
      provider: openai
      model: gpt-4o
      prompt: |
        Identify risks in this contract:
        ${input.contract_text}

        Return JSON: {
          "risks": [
            {
              "category": "financial" | "legal" | "operational",
              "severity": "high" | "medium" | "low",
              "description": string,
              "recommendation": string
            }
          ]
        }

  # Compare to standard template
  - name: compare-template
    operation: think
    config:
      prompt: |
        Compare contract to our standard template:

        Contract: ${input.contract_text}
        Template: ${component.standard-contract@v1.0.0}

        Identify deviations.

  # Legal review for high-risk
  - name: legal-review
    condition: ${JSON.parse(identify-risks.output).risks.some(r => r.severity === 'high')}
    agent: hitl
    inputs:
      data:
        terms: ${extract-terms.output}
        risks: ${identify-risks.output}
        deviations: ${compare-template.output}
      prompt: "Contract has high-risk clauses - legal review required"
      approvers: [legal@example.com]

ensemble: process-form

agents:
  # Extract form fields
  - name: extract
    operation: think
    config:
      provider: openai
      model: gpt-4o
      image: ${input.form_image}
      prompt: |
        Extract all form fields and values.
        Return as JSON object.

  # Validate completeness
  - name: check-complete
    operation: code
    config:
      code: |
        const form = JSON.parse(${extract.output});
        const required = ${input.required_fields};

        const missing = required.filter(field => !form[field]);

        return {
          complete: missing.length === 0,
          missing_fields: missing
        };

  # Request missing info
  - name: request-info
    condition: ${!check-complete.output.complete}
    operation: email
    config:
      to: ${input.submitter_email}
      subject: "Form incomplete"
      body: |
        Please provide:
        ${check-complete.output.missing_fields.join('\n')}

  # Process complete forms
  - name: process
    condition: ${check-complete.output.complete}
    operation: storage
    config:
      type: d1
      query: INSERT INTO form_submissions ...

config:
  provider: openai
  model: gpt-4o  # Supports vision
  image: ${input.document_url}

prompt: |
  Extract data.
  Return JSON: { "field1": type, "field2": type }

- name: validate
  agent: validator
  inputs:
    schema: ${component.schema@v1.0.0}

- name: review
  condition: ${confidence < 0.8 || high_value}
  agent: hitl

- name: log
  operation: storage
  config:
    query: INSERT INTO processing_log ...