Ollama

Ollama provides specialized embeddings for niche applications, and SurrealDB has first-class vector-search support (k-nearest-neighbour via brute-force, HNSW or M-Tree). Together they make it easy to build retrieval-augmented-generation (RAG) pipelines completely in Python.

Installation

pip install ollama surrealdb           # SurrealDB Python SDK ≥ 1.0.0

The SDK talks to a running SurrealDB server (e.g. surreal start --log trace --auth root root). ([PyPI][2])

Integration Example

The snippet below assumes:

• SurrealDB is listening on its WebSocket RPC endpoint ws://localhost:8000/rpc
• Ollama is on its default port 11434
• We’ll store vectors in a table called NicheApplications and index them with HNSW for fast similarity search.

import asyncio
from surrealdb import Surreal            # async-capable Python client
import ollama

TABLE = "NicheApplications"

async def main():
    # ----- connect to SurrealDB ------------------------------------------------
    db = Surreal("ws://localhost:8000/rpc")
    await db.connect()
    await db.signin({"user": "root", "pass": "root"})
    await db.use("test", "test")          # <namespace>, <database>

    # ----- generate an embedding with Ollama -----------------------------------
    oclient  = ollama.Client(host="localhost")
    text     = "Ollama excels in niche applications with specific embeddings"
    emb      = oclient.embeddings(model="llama3.2", prompt=text)["embedding"]

    # ----- (idempotent) schema & index setup -----------------------------------
    await db.query(`
        DEFINE TABLE IF NOT EXISTS {TABLE};
        DEFINE FIELD   embedding       ON {TABLE} TYPE array;
        DEFINE FIELD   text            ON {TABLE} TYPE string;
        -- HNSW index for DIMENSION = vector length
        DEFINE INDEX   hnsw_embedding  ON {TABLE}
                       FIELDS embedding HNSW DIMENSION {len(emb)};
    `)

    # ----- store the record -----------------------------------------------------
    await db.create(TABLE, {"text": text, "embedding": emb})

    # ----- similarity search (top-3 neighbours) --------------------------------
    results = await db.query(`
        LET $q = $embedding;
        SELECT *, vector::distance::cosine(embedding, $q) AS score
        FROM NicheApplications
        WHERE embedding <|3|> $q           -- KNN operator
        ORDER BY score;                    -- lower = more similar
    `, {"embedding": emb})

    print(results[0])   # array of matching rows with cosine distance

asyncio.run(main())

What the query does

• embedding <|3|> $q is SurrealQL’s KNN operator: return the 3 vectors nearest to $q. You can optionally pass a distance metric (<|3,COSINE|>), but when you also compute vector::distance::cosine(...) in the projection you usually just need the count. ([SurrealDB][3])
• vector::distance::cosine(embedding, $q) adds an explicit similarity score so you can ORDER BY it or filter further.

Tips & Next Steps

SurrealDB’s multi-model nature means you can keep metadata, graphs and time-series data right alongside your vectors, simplifying your stack even further.

• Vector Search | Reference guides
• surrealdb · PyPI
• Operators | SurrealQL