Skip to main content

Retrieval Inventory

FSR Inventory

Below is a list of various retrieval techniques and resources grouped by which part of the stack they influence

  • Query Transformation - Augment, structure, or enhance your input query
    • Multi-Query - Generate additional questions/queries based on your original query to return more holistic documents
  • Index - Adjust your data structures and associations
    • Multi-Vector - In addition to your the normal embeddings of your documents it is sometimes helpful to have multiple alternative embeddings per document. These can include embeddings of a summary, hypothetical questions, or any other custom text
    • Parent Document Retriever - A version of multi-vector where large chunks are further split up into smaller (child) chunks
  • Retrieval Methods - The method in which you pull documents out of your knowledge base
    • Top-K Similarity Search - Select the top K similar documents that match your query from your vectorestore. This is the 'Hello World' example of retrieval.
    • Maximum Marginal Relevance (MMR) - Return similar but diverse documents. Great for when you'd like to remove redundancy in your context
  • Document Transform - Transform your documents before using them as context with your LLM
    • Contextual Compression - Extract contextually relevant information from your retrieved docs. Generally used to try and to increase signal:noise ratio