Retrievers explained

What is a Retriever

A Retriever in Gen AI Builder is a specialized Griptape Structure that defines how an AI Assistant retrieves relevant information from one or more Knowledge Bases at query time.

Retrievers form the core of Retrieval-Augmented Generation (RAG) pipelines:

• They take the user’s question or Assistant task as input.
• They run semantic and optional structured search across configured Knowledge Bases.
• They return the most relevant "chunks" or documents.
• The Assistant passes this retrieved content to the Large Language Model (LLM) as part of its input context.

In short: Retrievers connect your Assistant to your Knowledge Bases — grounding the AI in your organization’s content.

Why use Retrievers

• Ground AI responses: Prevent hallucination by anchoring responses in trusted, current content.
• Enable explainability: Retrieved content can be shown or cited in AI responses.
• Power multi-KB search: Assistants can target multiple Knowledge Bases via a single Retriever.
• Customize retrieval behavior: Max Tokens, metadata, and Griptape Structure logic enable fine-tuning.

Without Retrievers, Assistants can only rely on their base model knowledge — which is static and cannot reflect your organization's current content.

How do Retrievers work

Under the hood:

1. User asks a question.
2. Assistant sends this query to its assigned Retriever(s).
3. Retriever (Griptape Structure) runs:

• Vector search on unstructured content (using embeddings).
• Structured filtering (if Hybrid KB is used and filters are specified).
• Ranking and selection of relevant results.

4. Retriever returns results to the Assistant:

• Content is formatted as context for the LLM.
• Optionally, references and metadata can be passed along.

5. Assistant generates its response using the retrieved content.

This entire flow happens at query time — ensuring responses reflect the most current content in your Knowledge Bases.

The role of Griptape Structures

Retrievers are implemented as Griptape Structures — making them highly customizable.

The core structure typically defines:

• Query preprocessing logic (if needed)
• Search logic (vector search + filters)
• Result post-processing and formatting
• Optional metadata injection

This architecture allows organizations to:

• Implement custom ranking strategies.
• Apply domain-specific filters.
• Enrich retrieved results with contextual tags.

See Structures explained for more on Griptape Structures.

Tuning and configuration points

Knowledge Bases selection

• Select one or more Knowledge Bases to target.
• A Retriever can search across multiple KBs — useful for broad Assistants.

Max Tokens

• Controls the total size of content passed to the LLM.
• Tuning guidance:
• Too low → insufficient context for high-quality answers.
• Too high → risk of LLM truncating the prompt or blowing past token limits.
• Typical values: 1000–3000 depending on use case and LLM in use.

Metadata

• Optional JSON object.
• Use for:
• Internal tracking and versioning
• Configuring advanced search behavior (if supported by your Retriever Structure)
• Passing hints to Assistants

Filters (Hybrid KB)

• If targeting a Hybrid Knowledge Base, structured filters can be applied (e.g., region, product, price).
• The Retriever Structure must support applying these filters.

Patterns of use

Single-KB Retriever

• Simple pattern — target one Knowledge Base.
• Example: HR Policy Retriever → queries HR Policy KB.

Multi-KB Retriever

• Targets multiple Knowledge Bases.
• Useful when an Assistant covers broad domains.
• Example: Support Assistant Retriever → queries Product Docs KB, Pricing Policies KB, Support FAQs KB.

Hybrid KB Retriever

• Targets Hybrid Knowledge Base.
• Supports structured filtering + semantic search.
• Example: Product Catalog Retriever → allows user to search Product Catalog Hybrid KB with filters for Category, Price, etc.

Cross-domain Retriever

• Designed for Assistants that switch "personas" or tasks.
• Can be implemented as:
• Single Retriever with dynamic KB selection
• Multiple Retrievers assigned per task or Assistant persona

Common tuning scenarios

Related topics

• Retrievers in Gen AI Builder
• Create a Retriever
• Knowledge Bases explained
• Retrieval Augmented Generation (RAG)
• Embeddings explained
• Structures explained
• Griptape concepts
• Hybrid Manager: Using Gen AI Builder