Skip to main content

LlamaIndex Integration

vectlite.llamaindex.VectLiteVectorStore implements the LlamaIndex VectorStore protocol. Plug it into VectorStoreIndex, query engines, chat engines, and any LlamaIndex flow that takes a vector store.

pip install vectlite llama-index-core

llama-index-core is an optional dependency of vectlite — installing it unlocks the integration.

Quick start​

from vectlite.llamaindex import VectLiteVectorStore
from llama_index.core import StorageContext, VectorStoreIndex, Document

store = VectLiteVectorStore(path="my.vdb", dimension=1536)
storage_ctx = StorageContext.from_defaults(vector_store=store)

documents = [
    Document(text="Auth setup guide for SSO", metadata={"source": "docs"}),
    Document(text="Billing FAQ", metadata={"source": "faq"}),
]
index = VectorStoreIndex.from_documents(documents, storage_context=storage_ctx)

query_engine = index.as_query_engine()
response = query_engine.query("How do I configure SSO?")
print(response)

Supported methods​

MethodDescription
add(nodes)Insert LlamaIndex BaseNode objects with their embeddings.
delete(ref_doc_id)Remove all nodes belonging to a reference document.
query(query)Vector / metadata search, returns a VectorStoreQueryResult.

Persistence across runs​

Because vectlite persists to a .vdb file, you don't need to re-ingest on every run. Reopen the existing store and rebuild the index from it:

store = VectLiteVectorStore(path="my.vdb")
index = VectorStoreIndex.from_vector_store(store)
query_engine = index.as_query_engine()

Chat engine example​

chat_engine = index.as_chat_engine(chat_mode="condense_question")
response = chat_engine.chat("What does the docs say about SSO?")

Filtering​

Pass MetadataFilters through the query engine — they map to vectlite metadata filters:

from llama_index.core.vector_stores import MetadataFilter, MetadataFilters

filters = MetadataFilters(filters=[
    MetadataFilter(key="source", value="docs")
])
query_engine = index.as_query_engine(filters=filters)

When to pick vectlite over other LlamaIndex stores​

  • No external dependency — a single .vdb file ships with your app. No Postgres + pgvector, no Pinecone account.
  • Crash-safe persistence — the WAL means an OS crash mid-ingest doesn't corrupt the index.
  • Built-in hybrid search — combine LlamaIndex's structured retrieval with vectlite's sparse + dense fusion.

Combining with vectlite-native features​

The underlying database is reachable via store.db — use it for TTL, payload indexes, quantization, or any feature outside LlamaIndex's protocol:

store.db.create_index("source", "keyword")
store.db.enable_quantization("scalar")