langchain-chassisdb

A LangChain vector store on Chassis: semantic search over your documents, kept in two files, inside your own process. No server to run, nothing to connect to.

pip install langchain-chassisdb

from langchain_chassisdb import ChassisVectorStore
from langchain_openai import OpenAIEmbeddings  # or any LangChain embeddings

store = ChassisVectorStore("docs.chassis", embedding=OpenAIEmbeddings())
store.add_texts(
    ["Refunds take five working days.", "We ship to the EU and the UK."],
    metadatas=[{"topic": "billing"}, {"topic": "shipping"}],
)

retriever = store.as_retriever(search_kwargs={"k": 4, "filter": {"topic": "billing"}})
retriever.invoke("how long does a refund take?")

• Files: docs.chassis holds the vectors and docs.db the documents and metadata. Open the same path again, with the same embeddings, to pick up where you left off. Everything added or deleted is saved when the call returns.
• Ids: give your own with ids=; adding a document with an id that is stored replaces it. get_by_ids and delete take them.
• Filters: filter={"key": value, ...} keeps documents whose metadata has every key with that value.
• Scores: similarity_search_with_score gives distances (smaller is closer); similarity_search_with_relevance_scores gives 0 to 1. max_marginal_relevance_search picks documents that are close to the query and different from each other.
• Options: metric="cosine" (the default) or "euclidean"; precision="half" keeps vectors in about half the space, with the same recall on the embeddings measured. They're fixed when the store is created: open it again with the same ones, or an OptionsMismatchError says which to pass. A read_only store takes them from the file.
• More processes: one process writes a store at a time. Others open it with read_only=True, take no lock, and see what the writer last saved.

It passes LangChain's standard tests for vector stores, sync and async.