AI Primitive Nodes

AI Primitive nodes are the building blocks of intelligent pipelines. They handle prompt rendering, lens execution, agent delegation, output parsing, vector embedding, RAG retrieval, evaluation, memory, and multi-step AI chains.

Node	Type	Output
Prompt Template	prompt_template	text
Lens Execute	lens_execute	lens_result
Agent Execute	agent_execute	agent_result
Output Parser	output_parser	json
Embedding	embedding	embedding
RAG Retriever	rag_retrieval	document[]
Vector Search	vector_search	document[]
Judge / Eval	judge_evaluator	json
Memory Read	memory_read	document[]
Memory Write	memory_write	json
Chain	chain	json
Summarizer	summarizer	text
Classifier	classifier	json
Translator	translator	text
Image Analyze	image_analyze	json
Audio Transcribe	audio_transcribe	text
Video Analyze	video_analyze	json

---

Prompt Template

Type: prompt_template · Category: AI Primitive

Render a prompt from variables and upstream data. Use before lens_execute or lens nodes to construct the input text.

Required Config

Field	Type	Description
template	template	Prompt template using {{variable}} syntax.

Optional Config

Field	Type	Description
variables	json	Explicit variable mappings: { varName: "$.sourcePath" }.

Example

{
  "template": "Summarize these arena results for {{audience}}: {{results}}",
  "variables": { "audience": "founders", "results": "$.results" }
}

Expected output: { "prompt": "Summarize these arena results for founders: ..." }

Downstream: → lens_execute with { "prompt": "$.prompt" }

Related Nodes

Lens Execute · Lens Node · Output Parser

---

Lens Execute

Type: lens_execute · Category: AI Primitive

Execute a selected LenserFight lens as a utility node (unlike the top-level Lens node, this is a catalog-driven step with an explicit lensId).

Required Config

Field	Type	Description
lensId	string	Lens id to execute.
model_id	string	Model key (e.g. openai:gpt-4.1-mini).

Optional Config

Field	Type	Description
param_overrides	json	Lens parameter overrides.

Example

{
  "lensId": "lens_weekly_digest",
  "model_id": "openai:gpt-4.1-mini",
  "param_overrides": { "tone": "crisp" }
}

Expected output: { "lensResult": { "text": "Weekly digest...", "modelId": "openai:gpt-4.1-mini" } }

Downstream: → email_send with { "body": "$.lensResult.text" }

Related Nodes

Lens Node · Prompt Template · Judge / Eval

---

Agent Execute

Type: agent_execute · Category: AI Primitive

Delegate work to a configured LenserFight agent. The agent runs autonomously and returns a structured result.

Required Config

Field	Type	Description
agentId	string	Agent id to invoke.
task	template	Task prompt with {{variable}} interpolation.

Optional Config

Field	Type	Default	Options
delegationPolicy	select	auto	auto · approval_required · forbidden

Example

{
  "agentId": "agent_pr_reviewer",
  "task": "Review PR {{prNumber}} for security and tests.",
  "delegationPolicy": "approval_required"
}

Expected output: { "agentResult": { "status": "completed", "summary": "Found 2 issues", "artifacts": [] } }

Downstream: → github_pr_review with { "reviewBody": "$.agentResult.summary" }

---

Output Parser

Type: output_parser · Category: AI Primitive

Parse model text into strict JSON fields. Use after a Lens or Lens Execute node when the model returns JSON-embedded in text.

Optional Config

Field	Type	Default	Description
schema	json	—	Expected output schema for type coercion.
strict	boolean	true	Fail when parsing is incomplete.

Example

{
  "schema": { "score": "number", "reasoning": "string", "winner": "string" },
  "strict": true
}

Expected output: { "score": 0.86, "reasoning": "Candidate A cites more evidence.", "winner": "candidate_a" }

Downstream: → judge_evaluator

Valid Connections

→ judge_evaluator, score_aggregator, leaderboard_update

Invalid Connections

✗ Cannot receive structured json input directly — it expects model-generated text.

Related Nodes

Lens Execute · Judge / Eval

---

Embedding

Type: embedding · Category: AI Primitive

Convert text or documents into embedding vectors with metadata. Use for building vector indexes, semantic similarity search, and RAG pipelines.

Inputs

Name	Type	Required
content	text	No
documents	document[]	No

At least one of content or documents is required.

Outputs

Name	Type	Shape
embedding	embedding	{ vector: number[], dimensions: number, metadata: json }

Required Config

Field	Type	Description
provider	select	openai · google · mistral
model	string	Embedding model (e.g. text-embedding-3-small).

Optional Config

Field	Type	Default
inputPath	string	$.text
chunkSize	number	1000
dimensions	number	1536
metadataFields	string[]	—

Example

{
  "provider": "openai",
  "model": "text-embedding-3-small",
  "inputPath": "$.documents[*].pageContent",
  "chunkSize": 1000,
  "dimensions": 1536,
  "metadataFields": ["battleId", "contenderId"],
  "retry": { "attempts": 3, "backoffMs": 1500 }
}

Scenario: Embed battle reports for later RAG retrieval.

Downstream: → vector_search with { "vector": "$.embedding.vector" }

Valid Connections

→ vector_search (direct vector lookup)

→ rag_retrieval (query embedding for retrieval)

Related Nodes

Text Splitter · RAG Retriever · Vector Search

---

RAG Retriever

Type: rag_retrieval · Category: AI Primitive

Retrieve scored documents from a vector source for a query. Use in question-answering, search, and context-augmented generation pipelines.

Inputs

Name	Type	Required
query	text	Yes
source	embedding	No

Outputs

Name	Type	Description
documents	document[]	Retrieved documents with scores.

Required Config

Field	Type	Description
vectorStore	string	Vector source (e.g. supabase:workflow_documents).
queryPath	string	Query mapping (e.g. $.rootInputs.query).

Optional Config

Field	Type	Default
topK	number	5
similarityThreshold	number	0.72
filters	json	—
rerank	boolean	true

Example

{
  "vectorStore": "supabase:workflow_documents",
  "queryPath": "$.rootInputs.query",
  "topK": 6,
  "similarityThreshold": 0.74,
  "filters": { "workspaceId": "{{workspace.id}}" },
  "rerank": true
}

Scenario: Retrieve knowledge base passages before answering a user question.

Downstream: → prompt_template with { "context": "$.documents[*].pageContent" }

Valid Connections

→ prompt_template, summarizer, judge_evaluator

Related Nodes

Embedding · Vector Search · Prompt Template

---

Vector Search

Type: vector_search · Category: AI Primitive

Search vectors directly using an embedding input.

Inputs

Name	Type	Required
embedding	embedding	Yes

Required Config

Field	Type	Description
vectorStore	string	Vector index name.

Optional Config

Field	Type	Default
topK	number	5

Example

{
  "vectorStore": "supabase:workflow_documents",
  "topK": 5,
  "filters": { "battleId": "$.metadata.battleId" }
}

Downstream: → summarizer

Related Nodes

Embedding · RAG Retriever

---

Judge / Eval

Type: judge_evaluator · Category: AI Primitive

Evaluate candidates against a rubric and emit structured scoring. The core node for AI-powered evaluation in battle pipelines.

Inputs

Name	Type	Required
candidates	json	Yes
rubric	text	No

Outputs

Name	Type	Shape
evaluation	json	{ score: number, reasoning: text, winner: text, confidence: number }

Required Config

Field	Type	Description
rubric	template	Evaluation rubric prompt.
scoringScale	string	Scale descriptor (e.g. 0-100).

Optional Config

Field	Type	Default
judgeModel	string	—
tieBreakRule	string	—
confidenceThreshold	number	0.7

Example

{
  "rubric": "Score correctness, source coverage, and actionability. Return JSON.",
  "scoringScale": "0-100",
  "candidateMappings": { "a": "$.answerA", "b": "$.answerB" },
  "judgeModel": "anthropic:claude-3-7-sonnet",
  "tieBreakRule": "prefer_higher_source_coverage",
  "confidenceThreshold": 0.76
}

Expected output: { "score": 91, "reasoning": "Answer A is more complete.", "winner": "answerA", "confidence": 0.84 }

Downstream: → score_aggregator with { "score": "$.score", "winner": "$.winner" }

Valid Connections

→ score_aggregator, leaderboard_update, slack_notify

Related Nodes

Battle Execute · Output Parser · Score Aggregator

---

Memory Read

Type: memory_read · Category: AI Primitive

Read conversation or workflow memory entries from a named namespace.

Required Config

Field	Type	Description
memoryKey	string	Memory namespace (e.g. workspace:arena-digests).

Optional Config

Field	Type	Default
limit	number	10

Example

{ "memoryKey": "workspace:arena-digests", "limit": 8, "query": "$.query" }

Expected output: { "documents": [{ "pageContent": "Last digest favored concise answers." }] }

Downstream: → prompt_template with { "memory": "$.documents" }

Related Nodes

Memory Write · RAG Retriever

---

Memory Write

Type: memory_write · Category: AI Primitive

Write durable workflow memory to a named namespace.

Required Config

Field	Type	Description
memoryKey	string	Memory namespace.
contentPath	string	Content mapping (e.g. $.summary).

Optional Config

Field	Type	Default	Options
policy	select	on_success	on_success · checkpoint

Example

{
  "memoryKey": "workspace:arena-digests",
  "contentPath": "$.summary",
  "policy": "checkpoint",
  "metadata": { "source": "weekly_digest" }
}

Downstream: → logger

Related Nodes

Memory Read

---

Chain

Type: chain · Category: AI Primitive

Run a configured AI chain of prompt, model, and parser steps sequentially.

Required Config

Field	Type	Description
steps	json	Chain step definitions. Each step: { type, ...stepConfig }.

Example

{
  "steps": [
    { "type": "prompt_template", "template": "Summarize {{input}}" },
    { "type": "lens_execute", "model_id": "openai:gpt-4.1-mini" },
    { "type": "output_parser", "schema": { "summary": "string" } }
  ]
}

Expected output: { "summary": "PR changes add catalog-backed workflow nodes." }

Downstream: → slack_notify with { "text": "$.summary" }

---

Summarizer

Type: summarizer · Category: AI Primitive

Summarize text or documents using a selected model.

Required Config

Field	Type	Description
model	string	Model key.

Example

{ "provider": "openai", "model": "gpt-4.1-mini", "inputPath": "$" }

Downstream: → email_send

Related Nodes

Prompt Template · Lens Execute · Text Splitter

---

Classifier

Type: classifier · Category: AI Primitive

Classify text into configured labels with confidence scores.

Required Config

Field	Type	Description
model	string	Model key.

Example

{ "provider": "openai", "model": "gpt-4.1-mini", "inputPath": "$" }

Expected output: { "label": "high_risk", "confidence": 0.87 }

Downstream: → switch

---

Translator

Type: translator · Category: AI Primitive

Translate text into a target language.

Required Config

Field	Type	Description
model	string	Model key.

Example

{ "provider": "openai", "model": "gpt-4.1-mini", "instructions": "Create translation for localized founder note." }

Downstream: → notion_write

---

Image Analyze

Type: image_analyze · Category: AI Primitive

Analyze an image and return structured observations (labels, objects, text, sentiment).

Inputs

Name	Type	Required
image	image	Yes

Example

{ "provider": "openai", "model": "gpt-4.1-mini" }

Downstream: → judge_evaluator for media moderation review.

---

Audio Transcribe

Type: audio_transcribe · Category: AI Primitive

Transcribe audio into text with timestamps.

Inputs

Name	Type	Required
audio	audio	Yes

Example

{ "provider": "openai", "model": "gpt-4o-transcribe" }

Downstream: → summarizer for meeting summary pipeline.

---

Video Analyze

Type: video_analyze · Category: AI Primitive

Analyze video frames and transcript into structured notes.

Inputs

Name	Type	Required
video	video	Yes

Example

{ "provider": "openai", "model": "gpt-4.1-mini" }

Downstream: → summarizer for battle replay review.

---

See also: Node Catalog Index · Battle Nodes · Data Nodes · Embedding and RAG concepts · Workflow Studio