Class Introduction

Function

Configures data structure of the similarity cache.

Prototype

from mx_rag.cache import SimilarityCacheConfig

SimilarityCacheConfig(vector_config, cache_config, emb_config, similarity_config, retrieval_top_k, clean_size, **kwargs)

Parameters

Parameter	Data Type	Required/Optional	Description
vector_config	Dict[str, Any]	Required	Used to configure a vector database. For details about the configuration, see Table 1. If the Faiss database is configured, load_local_index will be overwritten by data_save_folder and auto_save will be False in MindFAISS. The length of the character string in the configuration cannot exceed 1024, and the length of the iterable sequence in the dictionary cannot exceed 1024, the dictionary length cannot exceed 1024, and the number of nested dictionary layers cannot exceed 2.
cache_config	String	Required	Used to configure a scalar database. The value can only be sqlite.
emb_config	Dict[str, Any]	Required	Used to configure an embedding model. For details, see Table 2. The dictionary length cannot exceed 1024 characters. The string length in the dictionary cannot exceed 1024 characters. The number of nested dictionary layers cannot exceed 1.
similarity_config	Dict[str, Any]	Required	Used to configure a similarity calculation model. The dictionary length cannot exceed 1024 characters. The string length in the dictionary cannot exceed 1024 characters. The number of nested dictionary layers cannot exceed 1. For details, see Table 3.
retrieval_top_k	Integer	Optional	Top k value during similarity retrieval. The default value is 1. The value range is (0, 1000].
clean_size	Integer	Optional	Number of aged cache data records when the number of added cache data records exceeds the value of cache_size. The default value is 1. The value range is (0, cache_size].
**kwargs	Any	Required	For details, see CacheConfig.

This API uses the pickle module, which may be attacked by maliciously constructed data during unpickle. Ensure that data_save_folder is securely stored and only trusted data can be loaded.
The values of vector_config and cache_config must be None or not None at the same time. If both vector_config and cache_config are None, the function is the same as that of memory cache.
For the SQlite database, the size of the flushed file cannot exceed 30 GB. For a vector database, the size of the flushed file cannot exceed 20 GB.

**Table 1** vector_config
Parameter	Data Type	Required/Optional	Description
**kwargs	Dict[str, Any]	Required	For details, see create_storage.
top_k	Integer	Optional	Number of top k records during similarity search. The default value is 5.
vector_save_file	String	Required	Flushing path. When vector_type is set to npu_faiss_db, this parameter overwrites the load_local_index parameter in MindFAISS. For milvus_db, this parameter does not take effect.

**Table 2** emb_config
Parameter	Data Type	Required/Optional	Description
x_dim	Integer	Optional	Number of dimensions of the embedding model. The default value is 0.
skip_emb	Bool	Optional	Whether to skip embedding. The default value is False.
**kwargs	Dict[str, Any]	Required	For details, see create_embedding.

**Table 3** similarity_config
Parameter	Data Type	Required/Optional	Description
score_min	Float	Optional	Minimum value of the similarity calculation range. The default value is 0.0. The value range is [0.0, 100.0].
score_max	Float	Optional	Maximum value of the similarity calculation range. The default value is 1. The value range is [1.0, 100.0]. The value of score_max must be greater than or equal to that of score_min.
reverse	Bool	Optional	Relationship between the similarity score and similarity. The default value is False. False: A higher similarity score indicates a higher similarity. True: A higher similarity score indicates a lower similarity.
**kwargs	Dict[str, Any]	Required	For details, see create_reranker.

Example

Example 1: faiss_npu + local_embedding + local_reranker

from mx_rag.cache import SimilarityCacheConfig
from mx_rag.cache import MxRAGCache
dim = 1024
dev = 1

similarity_config = SimilarityCacheConfig(
        vector_config={
            "vector_type": "npu_faiss_db",
            "x_dim": dim,
            "devs": [dev],
           
        },
        cache_config="sqlite",
        emb_config={
            "embedding_type": "local_text_embedding",
            "x_dim": dim,
            "model_path": "path_to_embedding_model", # Embedding model path
            "dev_id": dev
        },
        similarity_config={
            "similarity_type": "local_reranker",
            "model_path": "path_to_reranker_model",  # Reranker path
            "dev_id": dev
        },
        retrieval_top_k=1,
        cache_size=1000,
        clean_size=20,
        similarity_threshold=0.86,
        data_save_folder="path_to_cache_save_folder", # Flushing path
        disable_report=True
    )
similarity_cache = MxRAGCache("similarity_cache", similarity_config)

Example 2: milvus_db + tei_embedding + tei_reranker

import getpass
from paddle.base import libpaddle
from mx_rag.cache import SimilarityCacheConfig
from mx_rag.cache import EvictPolicy
from mx_rag.cache import MxRAGCache
from mx_rag.utils import ClientParam
from pymilvus import MilvusClient
dim = 1024

client = MilvusClient("https://x.x.x.x:port", user="xxx", password=getpass.getpass(), secure=True,   client_pem_path="path_to/client.pem",   client_key_path="path_to/client.key",   ca_pem_path="path_to/ca.pem",   server_name="localhost")
similarity_config = SimilarityCacheConfig(
    vector_config={
        "client": client,
        "vector_type": "milvus_db",
        "x_dim": dim,
       
        "collection_name": "mxrag_cache_123",  # Label of milvus_db
        "param": None
    },
    cache_config="sqlite",
    emb_config={
        "embedding_type": "tei_embedding",
        "url": "https://<ip>:<port>/embed",  # IP address and listening port of the tei_embedding service
        "client_param": ClientParam(ca_file="/path/to/ca.crt")
    },
    similarity_config={
        "similarity_type": "tei_reranker",
        "url": "https://<ip>:<port>/rerank",  # IP address and listening port of the tei_reranker service
        "client_param": ClientParam(ca_file="/path/to/ca.crt")
    },
    retrieval_top_k=1,
    cache_size=100,
    auto_flush=100,
    similarity_threshold=0.70,
    data_save_folder="path_to_cache_save_folder",
    disable_report=True,
    eviction_policy=EvictPolicy.FIFO
)
similarity_cache = MxRAGCache("similarity_cache", similarity_config)

Parent topic: SimilarityCacheConfig