deepSearch

Runs a multi-agent research workflow over one or more RAG datasets. Builds a LangGraph state machine composed of a researcher agent, a tools node (wrapping vectorSearch and rerank), an advisor agent that scores research quality and identifies evidence gaps, and a synthesize node that extracts structured facts. The graph iterates — researcher → tools → advisor → researcher — until the advisor scores quality ≥ 0.9 with no evidence gaps, or max_research_iterations is reached.

Dataset resolution: if collection_id is provided and dataset_ids is not, the node calls getCollection to resolve dataset IDs. If neither is provided, a ValueError is raised.

Parameters

Param	Type	Required	Default	Description
query	string	Yes	—	Research question. Minimum length: 1
dataset_ids	integer[]?	No	null	Dataset IDs to search. Takes priority over collection_id
collection_id	integer?	No	null	Collection to resolve dataset IDs from, when dataset_ids is not provided
max_research_iterations	integer	No	3	Maximum iteration cycles. Range: 1–10
max_concurrent_agents	integer	No	3	Maximum concurrent research agents. Range: 1–10
allow_clarification	boolean	No	false	Ask clarifying questions before research
max_tool_calls_per_iteration	integer	No	10	Maximum tool calls per iteration. Range: 1–30
research_model	string	No	"gpt-4o-mini"	LLM model for researcher and advisor agents
research_provider	string?	No	null	LLM provider. Auto-detected from model when not set
summarization_model	string?	No	null	LLM model for summarization. Falls back to research_model
final_report_model	string?	No	null	LLM model for the final report. Falls back to research_model
search_k	integer	No	10	Documents retrieved per RAG search call. Range: 1–50
embedding_model	string	No	"BAAI/bge-m3"	Embedding model for vector search
embedding_provider	string	No	"runpod"	Embedding provider
max_documents_per_agent	integer	No	20	Maximum documents each agent can analyze. Range: 1–100
report_format	string	No	"detailed"	One of: summary, detailed, comprehensive

Output

Field	Type	Description
extracted_facts	ExtractedFact[]	Structured facts extracted by the synthesis LLM
citations	Citation[]	Unique citations deduplicated by DOI
research_summary	string	Free-text summary produced by the synthesis LLM
evidence_gaps	string[]	Gaps in evidence identified by the advisor
total_documents_analyzed	integer	Count of unique documents analyzed across all iterations
total_iterations	integer	Number of researcher→tools→advisor cycles completed
research_quality_score	float	Final quality score assigned by the advisor (0.0–1.0)
execution_time_seconds	float	Wall-clock time in seconds

Each ExtractedFact:

Field	Type	Description
fact	string	Extracted fact
confidence_score	float	Confidence (0.0–1.0)
supporting_citations	integer[]	Citation id values supporting this fact
relevance_score	float	Relevance to the query (0.0–1.0)

Each Citation:

Field	Type	Description
id	integer	Sequential citation ID
title	string	Paper title
doi	string	DOI
subject_area	string	Subject area or field

Example

{
  "id": "research",
  "type": "deepSearch",
  "data": {
    "label": "Deep Search",
    "isExecuted": false,
    "handles": ["inputs", "outputs"],
    "schema": {},
    "params": {
      "query": { "value": "{{ $input.question }}", "isExpression": true, "isAttachedToInputNode": false },
      "collection_id": { "value": 42, "isExpression": false, "isAttachedToInputNode": false },
      "max_research_iterations": { "value": 4, "isExpression": false, "isAttachedToInputNode": false },
      "research_model": { "value": "gpt-4o-mini", "isExpression": false, "isAttachedToInputNode": false },
      "search_k": { "value": 25, "isExpression": false, "isAttachedToInputNode": false },
      "report_format": { "value": "detailed", "isExpression": false, "isAttachedToInputNode": false }
    },
    "inputs": [], "outputs": [], "errors": []
  },
  "position": { "x": 0, "y": 0 },
  "isSelected": false,
  "isDragging": false
}