Agentic RAG Retriever

AI Toolkit agentic_rag_retriever Professional

Advanced multi-hop RAG retrieval with iterative AI-driven search refinement

Overview

The Agentic RAG Retriever uses AI to perform multiple search iterations, refining its queries based on previous results until it finds adequate information.

How It Works

  1. Generate Queries: AI analyzes your retriever prompt and creates optimized search queries
  2. Search: Executes the queries across your knowledge collections
  3. Evaluate Results: AI determines if the results are sufficient
  4. Refine or Return: If results are adequate, returns them. Otherwise, generates new queries and searches again (up to the configured maximum)

Output Variable

The search results are stored:

Results: {{agentic_rag.results[0].content}}
Count: {{agentic_rag.count}}

Configuration Fields

Retriever Prompt handlebars Required

Instructions for what information to retrieve.

Show details

Supports Handlebars syntax:

{{variable_name}}
Knowledge Collections knowledge_collections Required

Select one or more knowledge collections to search.

Show details

Select one or more knowledge collections to search. Collections are managed in the Knowledge section.

Search Settings search_settings

Configure search behavior and result filtering.

Show details

Configuration options:

  • Search type (hybrid, semantic, keyword)
  • Result limit (1-50)
  • Quality score threshold
  • Similarity threshold

Search types:

  • Hybrid: Combines semantic and keyword matching for best results
  • Semantic: Finds conceptually similar content using AI embeddings
  • Keyword: Traditional exact/fuzzy text matching
AI Configuration agent_settings

Configure the AI model and behavior.

Show details

Configuration options:

  • AI model selection
  • Temperature (0.0-2.0) - Higher = more creative, Lower = more consistent
  • Conversation context length - Amount of chat history to include
  • Personality instructions - Defines how the AI interacts

Temperature guide:

  • 0.0-0.5: Focused, deterministic, predictable
  • 0.5-1.0: Balanced between creativity and consistency
  • 1.0-2.0: Creative, varied, exploratory
Error Handling error_handling

Configure how the node handles errors.

Show details

Configuration options:

  • Continue on error: If enabled, flow continues through error handle. If disabled, flow terminates.
  • Retry attempts (0-3): Number of times to retry the operation before failing
  • Retry delay (100-1000ms): Time to wait between retry attempts

Use retries for transient errors like network issues. Use "continue on error" to implement fallback logic.

Output Variables

This node stores its output in a variable that can be referenced in subsequent nodes using the {{variable}} syntax.

{{variable}}.results array

Array of search result objects with content, relevance scores, source information, and collection metadata

{{variable}}.results[].content string

The text content of the search result chunk

{{variable}}.results[].score number

Relevance score for this result (0.0-1.0)

{{variable}}.results[].source string

Source document name or URL where this result originated

{{variable}}.results[].collection_id string

Unique identifier of the collection this result came from

{{variable}}.results[].collection_name string

Human-readable name of the collection this result came from

{{variable}}.count number

Total number of results returned after all search iterations

{{variable}}.timestamp string

ISO 8601 timestamp when the final results were returned

{{variable}}.inadequate boolean

Optional flag indicating results were returned but marked as inadequate by AI evaluation

{{variable}}.search_exhausted boolean

Optional flag indicating max search iterations were reached before finding adequate results

Connection Points

Input
Accepts incoming connections from other nodes
✓ Results Found output
AI-driven search found adequate results after iterative refinement
∅ No Results empty_results
AI-driven search completed but found no adequate results

More in AI Toolkit