Apache Solr Vector Store for LlamaIndex

A LlamaIndex VectorStore using Apache Solr as the backend.

Install

pip install llama-index
pip install llama-index-vector-stores-solr

Solr References

• Getting Started
• Deployment Guide
• Indexing Guide
• Query Guide
• KNN/Vector Search Support
• Docker Setup

Quickstart

Imports

from llama_index.vector_stores.solr import (
    ApacheSolrVectorStore,
    SyncSolrClient,
    AsyncSolrClient,
)
from llama_index.core import (
    VectorStoreIndex,
    StorageContext,
    Document,
    MockEmbedding,
)
import pytest
from llama_index.core import Settings
from llama_index.core.vector_stores.types import (
    VectorStoreQuery,
    VectorStoreQueryMode,
)
from llama_index.core.schema import NodeWithScore

Setup Vector Store

SOLR_COLLECTION_URL = "http://localhost:8983/solr/my_collection"  # assumes a solr collection is running here

sync_client = SyncSolrClient(base_url=SOLR_COLLECTION_URL)
async_client = AsyncSolrClient(base_url=SOLR_COLLECTION_URL)

vector_store = ApacheSolrVectorStore(
    sync_client=sync_client,
    async_client=async_client,
    nodeid_field="id",
    content_field="content_txt_en",  # store content in a text field searchable by BM25
    embedding_field="vector_field",  # dense vector field configured in Solr schema
    metadata_to_solr_field_mapping=[
        ("author", "author_s"),
    ],
    text_search_fields=["content_txt_en"],
)

Create Index and Query

# Dummy Documents
docs = [
    Document(
        text="Apache Solr integrates with LlamaIndex.",
        metadata={"author": "alice"},
    ),
    Document(
        text="Vector search lets you find semantically similar text.",
        metadata={"author": "bob"},
    ),
]

# Create the index
storage_context = StorageContext.from_defaults(vector_store=vector_store)
# NOTE: This will use the default embedding model set at Settings.embed_model
# Configure your own if you wish to do so.
# Alternatively this Vectorstore can be used with the IngestionPipeline as well.
index = VectorStoreIndex.from_documents(docs, storage_context=storage_context)

BM25 Search

• This is a naive implementation ideally create a retriever with BaseRetriever

def simple_bm25(vector_store, query_str="semantic search"):
    results = vector_store.query(
        VectorStoreQuery(
            mode=VectorStoreQueryMode.TEXT_SEARCH,
            query_str=query_str,
            sparse_top_k=2,
        )
    )
    return [
        NodeWithScore(node=node, score=score)
        for node, score in zip(
            results.nodes, results.similarities or [], strict=True
        )
    ]


bm25_search_results = simple_bm25(vector_store, query_str="semantic search")

Dense Vector Search

• This is a naive implementation ideally create a retriever with BaseRetriever

# Dense Vector Search
def simple_vector_search(vector_store, query_str="semantic search"):
    retriever = index.as_retriever(similarity_top_k=1)
    return retriever.retrieve("semantic search")


vector_search_results = simple_vector_search(vector_store, "semantic search")

Query Engine

query_engine = index.as_query_engine(similarity_top_k=2)
# NOTE: Will use default embedding and LLM model set at Settings
# Configure your own if you wish to do so.
res = query_engine.query("semantic search")