Search Knowledge Base

AI search over a knowledge base

Overview

Runs an agentic search over a BusyBot knowledge base. A search subagent greps, reads and semantically analyzes the knowledge base documents, records evidence, and returns findings (per-document extracted content with confidence), a written answer to the query, and a summary of how the search was conducted. Supports a fast mode (up to 20 iterations) and a deep-research mode (40 iterations across 8 phases).

Category: Data & Storage
Tool Name: search_knowledge_base
Version: 1

Appearance: Icon: lucide-ScanSearch | Color: #7C3AED

Node Type

Action — processes input items and produces output

Input / Output

Direction	Port(s)
Input	Input
Output	Output, Error

Credentials

This tool does not require any credentials.

Parameters

Parameter	Type	Required	Default	Description
Knowledge Base	string	Yes	—	The knowledge base ID or its exact name. Supports expressions like {{ $json.kbId }}.
Query	string	Yes	—	What to find, in natural language. A full question works far better than keywords. Supports expressions.
Mode	options	No	quick	How thoroughly to search.
				Options: quick (free-form search, up to 20 iterations — the engine agent chat uses), deep (40 iterations across 8 gated research phases; slow and expensive, best for background runs)
Max Iterations	number	No	20	Quick mode only — upper bound on search iterations. Deep mode always runs its 40-iteration phase plan. (shown when Mode is quick)
Output Mode	options	No	single	How to shape the output items.
				Options: single (one output item per input item, findings nested in an array), perFinding (fan out — one output item per finding)
Include Findings	boolean	No	true	Include per-document extracted content. Turn off to return only the answer and summary. (shown when Output Mode is single)
Include Search Steps	boolean	No	false	Include the step-by-step trace of what the search agent did (useful for debugging a disappointing answer).
Fail If Nothing Found	boolean	No	false	Treat a search that returns no findings as an item error instead of a normal result.
Max Concurrency	number	No	2	Searches running at once. Each is a full LLM agent loop — keep this low.

Output Data

One search per input item. Results are written to kbSearch on the output item; the rest of the item JSON passes through unchanged, and binary data is forwarded.

{
  "kbId": "kb_...", "kbName": "Product Manuals", "query": "...", "mode": "quick",
  "answer": "The direct answer synthesized from the findings",
  "searchSummary": "What was searched, which documents were examined, how the answer was reached",
  "findings": [{ "docId": "doc_...", "fileName": "manual.pdf", "relevantContent": "the actual extracted text", "confidence": "high" }],
  "findingCount": 3, "iterations": 6, "exitReason": "model",
  "documentsInKb": 42, "stale": false, "healthy": true, "durationMs": 18420
}

• exitReason: model (the agent finished on its own — the good case), cap (hit the iteration limit), stall (repeated itself and was cut short), phased (deep mode ran its full course).
• stale: true means documents changed since the knowledge base summary was regenerated; results may not reflect the latest content.
• findings is present in single output mode when Include Findings is on. In perFinding mode the node fans out instead: one output item per finding, each carrying finding and findingIndex with the answer and summary repeated. A search with no findings still emits one item, with finding: null.
• steps is added when Include Search Steps is on, and phasesVisited when Mode is deep.

Reference the result downstream by expression, e.g. {{ $json.kbSearch.answer }}.

Usage Examples

• search the knowledge base for the payment terms
• ask my knowledge base what the maintenance interval is
• do deep research across the knowledge base on data retention commitments
• find every document mentioning the vendor and quote the relevant lines

Example Configuration

Answer a question from a knowledge base:

{
  "type": "search_knowledge_base",
  "parameters": {
    "knowledgeBase": "Product Manuals",
    "query": "What is the recommended maintenance interval for the {{ $json.model }} pump?",
    "mode": "quick"
  }
}

Deep background research, fanned out one item per finding:

{
  "type": "search_knowledge_base",
  "parameters": {
    "knowledgeBase": "kb_1765432100000_ab12cd34",
    "query": "Every commitment we made about data retention, with the document it came from",
    "mode": "deep",
    "outputMode": "perFinding",
    "includeSteps": true
  }
}

Error Handling

Mode	Behavior
stop	Halts workflow on first error
continue	Skips failed items, passes successful ones through
errorPort	Routes failed items to Error output port

Tips

Searches a knowledge base with an AI research subagent and returns what it found: per-document findings with the actual extracted content and a confidence level, a written answer to the query, and a summary of how the search was run. Use quick mode for normal lookups and deep mode for exhaustive background research — deep runs 40 phased iterations (map, synonym expansion, semantic deep-read, cross-checking, gap analysis, verification, synthesis) and costs substantially more time and tokens. Point it at a knowledge base by ID or name and give it a natural-language question rather than keywords.

Choosing a mode

• quick — the default. Typical searches finish in 3–8 iterations. Use for lookups, enrichment inside a live pipeline, and anything a person is waiting on.
• deep — a scheduled/background research job over a large or high-stakes corpus: exhaustive enumeration, “is X mentioned anywhere”, contradiction hunting, or literature-style review. It always runs its phases and can take many minutes.

Behavior notes

• The knowledge base must be ready (ingested). Searching an empty or still-processing KB errors.
• The content bundle is downloaded once per knowledge base per node run and shared by every item targeting it, then deleted.
• Cost is real: llm_search sends document text to a model, and deep mode is designed to spend iterations. Prefer quick mode plus a narrow query.