Langchain

SurrealDB is a multi-model database that ships two built-in vector-search algorithms:

You expose either one with a single DEFINE INDEX … HNSW|MTREE statement in SurrealQL

Install the packages

pip install surrealdb langchain-community langchain-openai   # swap in any embedding provider you like

• surrealdb → Python SDK
• langchain-community → houses SurrealDBStore
• langchain-openai (or HF, Cohere, etc.) → embeddings

Quick start – create a vector store from text

from langchain_community.vectorstores.surrealdb import SurrealDBStore
from langchain_openai import OpenAIEmbeddings          # use any Embeddings class

emb = OpenAIEmbeddings()                               # or HF, Cohere, etc.

store = SurrealDBStore.from_texts(
    texts,                                             # list[str]
    embedding=emb,
    dburl="ws://localhost:8000/rpc",                   # SurrealDB RPC endpoint
    ns="langchain", db="docstore", collection="texts",
    db_user="root", db_pass="root",
)

Under the hood the helper will:

1. Create table texts (if it doesn’t exist).
2. Add two fields – text & embedding.
3. Add an HNSW index with the correct dimensionality.
4. Insert each text with its freshly generated embedding.

Using an existing collection

If you’ve already ingested vectors (e.g. via another app) just instantiate the store directly:

from langchain_community.vectorstores.surrealdb import SurrealDBStore
store = SurrealDBStore(
    embedding_function=emb,
    dburl="ws://localhost:8000/rpc",
    ns="langchain",
    db="docstore",
    collection="texts",
    db_user="root",
    db_pass="root",
)

Local-mode SurrealDB for testing

In-memory

surreal start --mem --user root --pass root

Connect with dburl="ws://localhost:8000/rpc" and everything lives in RAM only – perfect for unit tests.

On-disk (single file)

surreal start file://./surreal.db --user root --pass root

Vectors (including the HNSW graph) are persisted between runs.

On-prem / cloud

Spin SurrealDB in Docker, K8s, Nomad, Fly.io, Railway – connection string stays the same (ws://host:8000/rpc or http://…/rpc).

Similarity search

Dense vector search (default)

query = "How do I enable vector search in SurrealDB?"
docs = store.similarity_search(query, k=3)   # cosine by default

If your table has an HNSW index, LangChain will issue a query that looks like:

SELECT *, vector::distance::knn() AS score
FROM texts
WHERE embedding <|3,64|> $q_vec
ORDER BY score;

• <|K,EF|> → SurrealDB’s built-in K-NN operator (K=3, efSearch=64).
• vector::distance::knn() → pulls the pre-computed distance for free. ([SurrealDB][2])

Exact search (no index)

Omit DEFINE INDEX and SurrealDBStore will fall back to vector::distance::cosine() for full-accuracy ranking.

Hybrid search (manual)

SurrealDB doesn’t yet bundle a sparse index, but you can:

1. Full-text index a content field with SEARCH in SurrealQL.
2. Retrieve an FTS candidate set, then add a vector_filter to LangChain’s retriever to run a second vector pass for re-ranking.

Switching algorithms

If you prefer M-Tree for exact search at the index level:

DEFINE INDEX mt_texts ON texts
  FIELDS embedding
  MTREE DIMENSION 768
  DIST   COSINE;

LangChain code stays unchanged – only your DEFINE INDEX differs.

Next steps

• SurrealDB vector-search reference
• LangChain API docs for SurrealDBStore
• YouTube deep-dive: Integrating SurrealDB with LangChain

That’s it – you now have a fully-featured LangChain vector store powered by SurrealDB’s built-in HNSW / M-Tree indexes, no boilerplate required. 🚀