AI Configuration Reference

Every Setting
for Every AI Block.

Complete reference documentation for configuring AI blocks in RenderDraw workflows. Use this page when you need to look up a specific setting, understand the interaction between parameters, or configure fallback and error handling behavior.

Core Settings

Model & Provider
Configuration.

provider

Type: string enum

Values: anthropic, openai, gemini, or a configured custom endpoint

Selects the AI provider for this block. Each block can use a different provider. Credentials are configured once per workspace and referenced by provider name.

model

Type: string

Anthropic: current Sonnet, Haiku, and Opus models configured in your workspace

OpenAI: current multimodal, fast, and reasoning models configured in your workspace

system

Type: string (markdown or plain text)

The system prompt. Defines the block's role, task, output format, and constraints. Context injections (knowledgebase chunks, prior block outputs) are appended after the system prompt automatically, clearly delimited.

Include an explicit output format description in the system prompt even when you set an outputSchema. Models follow format instructions more consistently when they appear in both places.

maxTokens

Type: integer

Default: 4096   Max: 8192 (Sonnet/Claude Sonnet), 32768 (Opus)

Maximum tokens in the model's response. Does not affect the input context size. Set to the minimum needed for your expected output — it reduces latency and cost on steps that generate short outputs.

Sampling Parameters

Control Output
Determinism.

temperature

Range: 0.0 – 1.0   Default: 0.3

Controls randomness. Lower values produce more consistent, deterministic output. Higher values introduce more variation and creativity.

0.0Extraction, formatting, counting
0.2Structured analysis
0.5Proposal drafting
0.7–1.0Creative writing (rarely appropriate)

top_p

Range: 0.0 – 1.0   Default: 1.0

Nucleus sampling parameter. Limits the vocabulary pool to the top-p cumulative probability mass. Use either temperature or top_p — not both. Most workflow blocks use temperature; leave top_p at 1.0 unless you have a specific reason to change it.

seed

Type: integer (optional)

When set, the model attempts to produce the same output for identical inputs. Use for reproducible testing and debugging. Note: seed provides near-determinism, not guaranteed exact reproduction, especially across model updates. Not supported on all providers.

Knowledgebase Context Injection

Inject the Right
Knowledge at Runtime.

When a knowledgebase context source is configured on an AI block, RenderDraw runs a semantic search against the knowledgebase at block execution time and injects the top-K retrieved chunks into the model's context, clearly delimited from the system prompt.

topK

integer, default: 5

Number of knowledgebase chunks to retrieve. Higher values inject more context but consume more tokens. Recommended: 3–7. Use 10+ only when reasoning across many distinct source documents.

threshold

float 0.0–1.0, default: 0.65

Minimum cosine similarity score for a chunk to be included. Chunks below this threshold are discarded even if they are in the top-K results. Increase to 0.75–0.85 to inject only high-confidence matches; lower to 0.5 to cast a wider net.

includeSources

boolean, default: false

When true, each injected chunk includes the source document name and page number as a citation prefix. The model is instructed to reference these citations in its output. Useful for audit trails on RFP responses.

queryOverride

string (optional)

By default, the semantic query uses the block's input document. Set queryOverride to a specific text string or a reference to a prior block's output field to use a different query for knowledge retrieval.

Error Handling

Retry Logic, Fallbacks,
and Graceful Degradation.

Fallback Model

Configure a fallback model that activates when the primary model returns an error (rate limit, overload, timeout). Specify the same or a different provider. Example: fallback from a Sonnet-class model to a Haiku-class model if the primary model is rate-limited.

"fallbackModel": "selected-fast-fallback-model"

Retry Logic

Configure automatic retries on transient errors (network timeouts, 429 rate limit responses, 500 server errors). RenderDraw uses exponential backoff with jitter. Settings: max retries (default: 3), initial delay (default: 1s), max delay (default: 30s).

"retry": { "maxAttempts": 3, "backoffMs": 1000 }

Graceful Degradation

Configure what happens when all retries and fallbacks are exhausted. Options: fail the block (stops the workflow and creates a human-review gate), skip the block (passes empty output downstream), or use default output (returns a configured static fallback value).

"onFailure": "gate" | "skip" | "default"
Complete Example

Full AI Block Configuration
for RFP Analysis.

blocks[1] — claude-rfp-analysis.json
{
  "id": "rfp-analysis",
  "type": "ai",
  "provider": "anthropic",
  "model": "selected-sonnet-model",

  "system": "You are an RFP analyst for
  Acme Construction Equipment Co.
  Extract every requirement from the
  attached RFP. Return a JSON array.
  For each requirement:
    id: sequential R-001, R-002, ...
    section: section heading from RFP
    category: technical|commercial|
              compliance|schedule|other
    text: verbatim requirement text
    complianceRisk: low|medium|high
    sourceDoc: citation if KB provided",

  "temperature": 0.1,
  "maxTokens": 8192,
  "seed": null,

  "context": [
    {
      "type": "knowledgebase",
      "id": "rfp-library-2024",
      "topK": 5,
      "threshold": 0.72,
      "includeSources": true
    },
    {
      "type": "prior_block_output",
      "blockId": "document-classification",
      "fields": ["rfp_type", "jurisdiction"],
      "format": "json"
    }
  ],

  "outputSchema": {
    "requirements": [{
      "id": "string",
      "section": "string",
      "category": "string",
      "text": "string",
      "complianceRisk": "string",
      "sourceDoc": "string|null"
    }]
  },

  "fallbackModel": "selected-fast-fallback-model",
  "retry": {
    "maxAttempts": 3,
    "backoffMs": 1000,
    "maxBackoffMs": 30000
  },
  "onFailure": "gate"
}

Configure Your First
AI Block.

Open RenderDraw, drag an AI block, and follow this reference to configure it correctly for your workflow step.