Not every workflow step needs the same model. RenderDraw lets you configure a different AI provider, model, system prompt, and context injection for every block in your workflow. Complex reasoning on one block, fast extraction on the next, structured generation on the last.
Using the same AI model for every step wastes money on steps that need speed, and sacrifices quality on steps that need deep reasoning. The right architecture uses the best model for each job — and wires them together.
Claude Sonnet reads the submitted construction PDF. Its multimodal vision extracts dimensions, counts structural elements, and identifies material symbols. This is what Claude Sonnet does best.
Claude Sonnet reads the extracted data plus the full RFP document (200k context window). It cross-references requirements, identifies gaps, and drafts a structured compliance matrix. This is what Claude does best.
A fast, cheap model reformats the draft into the required template structure. No heavy reasoning required — just reliable instruction-following at scale. Haiku runs this in under 3 seconds.
Every AI block in RenderDraw has a configurable context window built from four components. You control exactly what each block sees — and what it does not.
{
model: "selected-long-context-model",
system: "You are an RFP analyst...",
context: [
{
type: "knowledgebase",
id: "rfp-library-2024",
topK: 5,
threshold: 0.72
},
{
type: "prior_block_output",
blockId: "drawing-analysis",
format: "json"
}
],
input: workflow.currentDocument,
maxTokens: 8192,
temperature: 0.2
}
Best for: long-document analysis (200k context), complex reasoning, nuanced drafting, multi-step instruction following. Powers RFP analysis, contract review, and compliance matrix generation.
Best for: vision and multimodal tasks, structured data extraction from images, fast text completion. Powers takeoff automation (image analysis of 2D drawings) and fast formatting steps.
When an AI block runs, it does not only see the current document. It also sees the most relevant chunks from your knowledgebase — your past proposals, product specs, pricing sheets, and technical documentation — retrieved via semantic search at the moment of execution.
This is what prevents hallucination. The model is not guessing. It is drawing on your real institutional knowledge, retrieved and injected into the context automatically.
Knowledgebases →context.query({
knowledgebase: "rfp-library-2024",
query: extractedRequirements,
topK: 5,
threshold: 0.72,
includeSource: true
})
5 relevant chunks retrieved from past proposals. Injecting 3,840 tokens of context into the selected AI block.
Open RenderDraw, add an AI block, and connect Claude, OpenAI, Gemini, or a custom model endpoint to your workflow in minutes.