Every AI block type in the RenderDraw workflow engine — how to configure each one, which model provider to use for which task, and how to chain AI blocks together for complex multi-step reasoning pipelines.
Per-block model selection. Every AI block in RenderDraw is independently configurable — select a different model for each block in the same workflow. Use a fast model for classification, a long-context model for complex RFP analysis, a vision-capable model for drawing tasks, and a low-cost model for simple field extraction. Right model for right task, controlled at the block level.
The text reasoning block is the most flexible AI block in the engine. It accepts an arbitrary system prompt, injects workflow data as context, and returns either freetext or structured JSON output. Use it anywhere you need an LLM to reason about your workflow data and make a decision or produce content.
{{field_name}} template syntaxYou are an estimating assistant for a
mechanical contractor. Analyze the
provided RFP scope description and
classify the project.
RFP Customer: {{customer_name}}
Project Type Context: {{project_type}}
Scope Description:
{{scope_text}}Output schema:
{
"classification": "cooling_tower",
"complexity": "high",
"bid_recommendation": "bid",
"reasoning": "..."
}The document analysis block is designed for processing multi-page PDF and Word documents — the kind of large, complex documents common in AEC and manufacturing: RFPs with 80+ pages, technical specifications, contract documents, material safety data sheets.
The block handles document chunking, parallel analysis, and result aggregation automatically. You define the output schema — the fields you want extracted — and the block drives the LLM through the document to find and return those fields.
Best model for this task: Claude Sonnet or Claude Opus for long-document analysis. Claude's 200K context window handles large RFPs without chunking in many cases. For very large documents (>500 pages), use chunking mode with Claude Sonnet for cost efficiency.
The vision analysis block processes images — scanned drawings, floor plans, P&ID diagrams, site photos, equipment photos — using vision-capable LLMs. It can count objects, identify types, read labels and text, measure relative positions, and detect changes between drawing revisions.
This block is central to the RenderDraw takeoff workflow — converting 2D construction drawings into structured quantity takeoff data without requiring CAD file conversion.
Best model for vision: Use a vision-capable model with strong spatial reasoning for counting, layout analysis, text reading, and label extraction. Configure different models for different vision tasks using per-block provider selection.
Always pair with human gate. Vision analysis accuracy on complex construction drawings varies by drawing quality, scale, and symbol library. Always route vision analysis results through a human review gate before using quantities in a bid or proposal.
The structured extraction block is optimized for pulling specific, well-defined fields from unstructured text — when you know exactly what shape of data you need but it lives in freetext form. Unlike the document analysis block (which handles full documents), the extraction block is designed for focused extraction from shorter text segments: email bodies, form responses, notes, short specification sections.
Define your output schema using a simple JSON schema builder — field name, data type (string, number, date, boolean, array), description, and optional validation constraints. The block uses function calling / tool use mode to force the model to return a structured object matching your schema exactly.
If the model cannot extract a value for a required field, it either retries with corrective prompting or marks the field as null with a low confidence score — letting you handle missing data explicitly in downstream logic.
// Example extraction schema
{
"project_name": "string (required)",
"bid_due_date": "date (required)",
"estimated_value": "number (optional)",
"project_address": "string (required)",
"project_type": "enum: [
renovation,
new_construction,
service
]",
"bonding_required": "boolean"
}Schema validation with auto-retry. If the model returns a malformed or incomplete object, the block automatically retries with a corrective prompt showing the error. Up to 3 retry attempts before the block fails with a structured error — allowing your workflow to route to a manual fallback path.
The summarization block condenses long documents, email threads, meeting transcripts, or arrays of data records into concise summaries at a configured target length and detail level.
When the input is an array of records rather than a single document — for example, 47 past project records retrieved from the knowledgebase — the summarization block uses a map-reduce pattern: summarize each item, then synthesize across all summaries.
This enables workflows that aggregate insights from large sets of historical data without hitting model context limits. Common for quarterly pipeline reports and competitive analysis workflows.
The draft generation block produces long-form content — full proposal sections, cover letters, RFI responses, technical narratives — by combining a rich system prompt with structured context data from earlier workflow blocks. It is the "write" companion to the "read" capability of document analysis and extraction.
The output is always intended to flow to a human review gate before delivery — draft generation accelerates the writing process, it does not replace judgment.
The quality of AI-drafted content scales directly with the richness of the context injected. A well-configured draft generation block pulls from:
Best model for drafting: Claude Opus for complex, high-value proposals requiring nuanced technical writing. Claude Sonnet for routine sections and faster turnaround. Claude Haiku for simple form fills and standardized section templates.
Extract RFP requirements
Pull relevant past projects
Write proposal sections
Estimator reviews draft
Every AI block lets you select the model provider and model version independently. Configure the system-wide default in Settings, then override per-block where a specific model excels for a particular task.
Point to any OpenAI-compatible API endpoint — self-hosted Llama, fine-tuned models on Azure AI, AWS Bedrock. Provide base URL + API key + model ID.
Controls randomness in model output. Default 0.2 for business workflows (near-deterministic, consistent results). Increase to 0.7–0.9 for creative writing tasks like cover letter generation where variation is desirable.
Cap the output length for cost control. Set conservatively for extraction tasks (short outputs needed); set generously for draft generation blocks. Truncated outputs trigger a block warning rather than silently returning incomplete content.
Configure a fallback model provider per block. If the primary provider is unavailable (rate limit, outage), the block automatically retries against the fallback provider. Prevents single-provider outages from halting critical workflows.
For Anthropic models, RenderDraw automatically uses prompt caching for system prompts and large context documents. Repeated workflow runs processing similar documents benefit from cache hits — reducing latency and cost by up to 80% on cached content.
Start with a document analysis block on an RFP. See structured data extracted in seconds. Build from there.