Skip to main content

Retrieval: Overview

Retrieval-Augmented Generation (RAG) is an important component for creating accurate, reliable AI applications. Teammately's approach automates and simplifies the entire RAG pipeline through our AI Agent, reducing hallucinations caused by noisy data and inefficient retrieval processes.

Traditional RAG implementations require manual configuration of multiple components which can be automated:

  • Document Processing
  • Contextual Chunking
  • Vector Indexing
  • Query Processing

The Agentic approach means you can simply provide your goal and data—our AI Agent handles the complex implementation details automatically.

Hallucination Prevention Focus​

Hallucinations in AI systems frequently stem from poor quality data and retrieval failures. Teammately addresses main root causes like Data Quality, Context Preservation and Retrieval Optimization.

Output Params & Format​

When setting up a Knowledgebook, you can configure how retrieval results are returned with these output parameters:

ParameterDescriptionDefault
Top KMaximum number of chunks to return from retrieval10
Min SimilarityMinimum similarity threshold for retrieved chunks (0.0-1.0)0.60
Output FormatFormat of returned results (Bullet List or Array of Text)Bullet List

Technical Implementation Details​

Teammately's retrieval system uses the following technical implementation:

  • Similarity Calculation: Uses Cosine similarity exclusively (Euclidean distance is not currently supported)
  • Vector Database: Implemented with pgvector and indexed by diskANN for efficient similarity search
  • Performance Optimization: Automatic balancing of recall and precision for optimal retrieval

Custom Implementation​

If you require alternative similarity metrics or custom vector indexing methods, please contact Teammately support for enterprise implementation options.

Integration with Generation Pipeline​

Teammately's RAG system integrates as Knowledgebooks with our Generation capabilities. This approach allows to deliver more accurate, reliable, and contextually appropriate AI responses.