Multiple AI Contexts

Claude Sonnet Sees the Drawing.
Claude Writes the Response.
Haiku Formats the Output.

RenderDraw lets every block in your workflow use a different AI provider, model, system prompt, and context. Configure each independently. Pass outputs between them. Each block is context-isolated — it only sees what you explicitly give it.

The Pattern

Vision → Reasoning
→ Formatting.

The most effective multi-model workflow pattern for RFP automation uses three blocks, each optimised for its specific task. This pattern is also used for RFI responses and complex takeoff workflows.

Block 1 — Claude Sonnet Vision

Reads the submitted PDF or drawing. Extracts dimensions, quantities, symbols, and tabular data. Returns structured JSON.

OpenAI

Block 2 — Claude Sonnet

Receives Claude Sonnet's extraction + full RFP document + knowledgebase context. Reasons across all three to draft a compliance matrix and response sections.

Anthropic

Block 3 — Claude Haiku

Receives Claude Sonnet's draft. Reformats it into the required proposal template structure. Fast and cheap — no heavy reasoning needed here.

Anthropic

Output

Structured proposal sections written to workbook + Word template populated for human review

Output
Context Isolation

Each Block Sees
Only What You Give It.

Blocks are context-isolated by default. Block 2 does not automatically see everything Block 1 saw. You explicitly wire the output of one block into the input context of the next. This is intentional — it prevents context window bloat and keeps each block's reasoning focused.

You control three things for each context pass:

  • What to pass — the full output, a specific field, or a summarized version
  • How to format it — inject as JSON, as markdown, or as inline text within the system prompt
  • Where to place it — in the system prompt, the user message, or as a separate context block

Why isolation matters: If Block 2 received Block 1's entire raw context (system prompt, retrieved knowledgebase chunks, full document), you would double or triple the token cost with no benefit. Block 2 only needs Block 1's output — the extracted data — not how it was extracted.

Cost optimization: A typical RFP workflow can use a vision-capable model for extraction, a long-context model for reasoning, and a fast model for formatting. Using one premium model for every step costs materially more than matching model class to task.

Configuration

How to Configure
Multiple Contexts.

1. Add the first AI block and select its provider

In the Workflow Editor, drag an AI Block onto the canvas. In the block's settings panel, select a provider and a vision-capable model. Set the system prompt to describe the vision extraction task. Enable Vision input and wire the workflow's input document to this block.

2. Add the second AI block with a different provider

Add another AI block. Select the provider and long-context model you want for reasoning. In the Context section, add a Prior Block Output context source and select Block 1. Also add a Knowledgebase context source and select your RFP library. Set topK to 5 and threshold to 0.72.

3. Configure context pass format

For the Block 1 → Block 2 context pass, choose Format: JSON so the reasoning block receives structured extraction data. Optionally select only specific output fields (e.g., extracted_items, drawing_dimensions) rather than the full output to reduce token cost.

4. Add the formatting block

Add a third AI block. Select a fast, low-cost model. Wire Block 2's output as context. Set the system prompt to describe the formatting template. Set Temperature: 0.0 — you want deterministic formatting, not creative variation. Lower temperature on formatting blocks reduces hallucinated structure.

Cost Optimization

Use Expensive Models
Only Where They Earn It.

Use Large Models For

  • RFP requirement extraction from complex 200-page documents
  • Ambiguous contract language interpretation
  • Cross-referencing multiple documents
  • Generating nuanced, compliant proposal text

Use Fast Models For

  • Reformatting into a fixed template
  • Classifying short text into categories
  • Extracting structured data from clean, simple inputs
  • Generating short summaries of known content

Use Vision Models For

  • Reading 2D construction drawings and plans
  • Counting symbols and annotations on scanned PDFs
  • Extracting tables from image-based documents
  • Identifying materials from product photographs

Build a Multi-Context Workflow
in 15 Minutes.

Connect Claude Sonnet vision, Claude Sonnet reasoning, and a fast formatting model in a single workflow. No infrastructure required.