A streamlined Python library that enhances LLM query performance through semantic caching, making responses faster and more cost-effective.

As AI applications gain traction, the costs and latency of using large language models (LLMs) can escalate. VectorCache addresses these issues by caching LLM responses based on semantic similarity, thereby reducing both costs and response times.

Semantic Vector Cache leverages the power of LLMs to provide two main advantages:

1. Faster Responses: By caching, it significantly reduces latency, offering quicker feedback to user queries.
2. Lower Costs: It minimizes the number of direct LLM requests, thereby saving on usage costs.

VectorCache, akin to a more nuanced Redis, enables efficient caching by recognizing not just exact matches but also semantically similar queries. This efficiency is particularly useful in domains where queries within a specific topic or field are frequent.

Prerequisites:

• Ensure Python version is 3.9 or higher (python --version).
• pip install vector-cache

NOTE: vector-cache comes with support for multiple vector stores and storage databases, which are optional installs. Read here for more information

• Install from source

  • For library installation issues, upgrade pip: python -m pip install -r requirements.txt.

• Initialise vector cache:

Initialization Parameters for Vector Cache

The VectorCache class can be initialized with the following parameters:

• embedding_model (BaseEmbedding): The embedding model used to convert text or other data into vector embeddings. This must be an instance of a class that inherits from BaseEmbedding.

  • db (CacheStorageInterface): The storage interface for caching and retrieving vectors. This must be an instance of a class that inherits from CacheStorageInterface.

  • vector_store (VectorStoreInterface): The vector store interface for storing and searching vectors. This must be an instance of a class that inherits from VectorStoreInterface.

  • initial_similarity_threshold (float, optional, default=0.99): The initial similarity threshold used to determine if a cached vector is considered a hit.

  • target_hit_rate (float, optional, default=0.8): The target hit rate for cache queries. This is used if use_adjustable_threshold is set to True to dynamically adjust the similarity threshold.

  • min_threshold (float, optional, default=0.7): The minimum allowed similarity threshold for adaptive thresholding.

  • max_threshold (float, optional, default=1.99): The maximum allowed similarity threshold for adaptive thresholding.

  • adjustment_rate (float, optional, default=0.01): The rate at which the similarity threshold is adjusted to achieve the target hit rate.

  • verbose (bool, optional, default=False): If set to True, enables verbose logging for debugging purposes.

  • use_adjustable_threshold (bool, optional, default=False): If set to True, enables dynamic adjustment of the similarity threshold based on the target hit rate.

• total_queries (int): The total number of queries made to the vector cache. This is used internally to track performance metrics.

  • cache_hits (int): The number of cache hits. This is used internally to track performance metrics and adjust the similarity threshold if use_adjustable_threshold is enabled.

  Example

  vector_cache = VectorCache(
      embedding_model=my_embedding_model,
      db=my_cache_storage,
      vector_store=my_vector_store,
      initial_similarity_threshold=0.95,
      target_hit_rate=0.85,
      min_threshold=0.75,
      max_threshold=1.0,
      adjustment_rate=0.02,
      verbose=True,
      use_adjustable_threshold=True
  )
  

• Refer to the examples folder for sample usage.

VectorCache is designed to work with any LLM provider. It includes modules for:

• Embedding Models: Facilitates similarity searches through various vector_cache.embedding APIs.
• Cache Storage: Stores LLM responses for future retrieval based on semantic matches.
• Vector Store: Identifies similar requests using the input request's embeddings.
• Cache Manager: Manages cache storage and vector store operations, including eviction policies.
• Similarity Evaluator: Determines the similarity between requests to ascertain cache matches.

** Features: **

• Adaptive Threshold (Alpha) : The vector cache can now incrementally adjust the similarity threshold till a desired hit rate is met. This should be used with caution as the high threshold can cause non-relevant matches to be returned.

VectorCache supports multiple vector store options for efficient similarity search. For detailed information on how to use different vector stores, please refer to our Vector Stores README.

Supported vector stores include:

• Redis
• Qdrant
• Deeplake
• ChromaDB
• pgvector
• Pinecone
• Milvus

We're continuously working on expanding our support for other popular vector stores. If you don't see your preferred vector store listed, check our documentation for the most up-to-date information or consider contributing to add support for it!

Interested in contributing to VectorCache? Check our contribution guidelines.