CamelAI

Camel is a Python framework for building multi-agent LLM systems. SurrealDB is a multi-model database with first-class vector search (HNSW, brute-force or M-Tree).

By adding a SurrealDB storage driver you can keep your agents’ knowledge base—documents, metadata and embeddings—in a single database.

Installation

pip install "camel[vector-databases]" surrealdb     # Camel + async SurrealDB SDK
surreal start --log trace --auth root root          # run SurrealDB locally

A minimal SurrealStorage driver

Camel’s VectorStorage interface just needs three methods: add, query and clear. Below is a lightweight implementation that writes to a table called AgentVectors with an HNSW index.

# surreal_storage.py
import asyncio, math
from typing import List
from camel.storages import VectorStorage, VectorRecord, VectorDBQuery, VectorDBQueryResult
from surrealdb import AsyncSurreal
from camel.types import VectorDistance

class SurrealStorage(VectorStorage):
    def __init__(
        self,
        *,
        url: str = "ws://localhost:8000/rpc",
        table: str = "AgentVectors",
        vector_dim: int = 768,
        distance: VectorDistance = VectorDistance.COSINE,
        namespace: str = "agents",
        database: str = "demo",
        user: str = "root",
        password: str = "root",
    ):
        super().__init__(distance)
        self.url = url
        self.table = table
        self.dim = vector_dim
        self.ns = namespace
        self.db = database
        self.user = user
        self.password = password

    async def _ensure_table(self):
        async with AsyncSurreal(self.url) as db:
            await db.signin({"username": self.user, "password": self.password})
            await db.use(self.ns, self.db)
            await db.query(
                """
                DEFINE TABLE $table SCHEMALESS;
                DEFINE FIELD payload    ON $table TYPE object;
                DEFINE FIELD embedding  ON $table TYPE array;
                DEFINE INDEX hnsw_idx   ON $table
                             FIELDS embedding HNSW DIMENSION $dim;
                """,
                {
                    "table": self.table,
                    "dim": self.dim
                }
            )

    # ---------- VectorStorage interface ---------------------------------
    async def add(self, records: List[VectorRecord]) -> None:
        await self._ensure_table()
        async with AsyncSurreal(self.url) as db:
            await db.signin({"username": self.user, "password": self.password})
            await db.use(self.ns, self.db)
            rows = [
                {"id": f"vec:{i}", "payload": r.payload, "embedding": r.vector}
                for i, r in enumerate(records)
            ]
            await db.create(self.table, rows)

    async def query(self, q: VectorDBQuery) -> List[VectorDBQueryResult]:
        await self._ensure_table()
        metric = {
            VectorDistance.COSINE: "cosine",
            VectorDistance.EUCLIDEAN: "euclidean",
            VectorDistance.DOT: "dot",
        }[self.distance]

        async with AsyncSurreal(self.url) as db:
            await db.signin({"username": self.user, "password": self.password})
            await db.use(self.ns, self.db)
            result = await db.query(
                """
                SELECT payload,
                       vector::distance::$metric(embedding, $vec) AS score
                FROM $table
                WHERE embedding <|$top_k,$metric|> $vec
                ORDER BY score;
                """,
                {
                    "table": self.table,
                    "vec": q.query_vector,
                    "top_k": q.top_k,
                    "metric": metric
                }
            )

            return [
                VectorDBQueryResult(
                    record=VectorRecord(vector=[], payload=row["payload"]),
                    similarity=1.0 - row["score"] if metric == "cosine" else -row["score"],
                )
                for row in result[0]["result"]
            ]

    async def clear(self) -> None:
        async with AsyncSurreal(self.url) as db:
            await db.signin({"username": self.user, "password": self.password})
            await db.use(self.ns, self.db)
            await db.query(
                "REMOVE TABLE $table;",
                {"table": self.table}
            )

Note

The driver uses SurrealDB’s Cosine, Euclidean or Dot-product distances by mapping Camel’s VectorDistance enum to SurrealQL functions.

Store and query vectors with Camel

import asyncio
from camel.storages import VectorRecord, VectorDBQuery
from camel.types import VectorDistance
from surreal_storage import SurrealStorage

storage = SurrealStorage(
    vector_dim=768,
    distance=VectorDistance.COSINE,
    user="root",  # Explicitly set credentials
    password="root"
)

async def demo():
    # ------------ write -----------------------------------------------
    await storage.add([
        VectorRecord(vector=[0.1]*768, payload={"title": "Alpha"}),
        VectorRecord(vector=[0.2]*768, payload={"title": "Beta"}),
    ])

    # ------------ similarity search -----------------------------------
    hits = await storage.query(VectorDBQuery(
        query_vector=[0.11]*768,
        top_k=5,
    ))
    for h in hits:
        print(h.record.payload, h.similarity)

    await storage.clear()

asyncio.run(demo())

Plug into Camel’s retrievers

from camel.embeddings import OpenAIEmbedding
from camel.retrievers import VectorRetriever

# SurrealDB storage with explicit configuration
surreal_storage = SurrealStorage(
    vector_dim=1536,
    user="root",
    password="root"
)

# Retriever with any embedding model Camel supports
retriever = VectorRetriever(embedding_model=OpenAIEmbedding())

# Process (embed + store) a piece of content
retriever.process("https://example.com/LLM-primer.pdf", surreal_storage)

# Query
results = retriever.query("What is an LLM?", surreal_storage)
print(results)

Auto-retriever convenience

from camel.retrievers import AutoRetriever
from camel.types import StorageType

auto = AutoRetriever(
    storage_type=StorageType.CUSTOM,         # tell Camel we use a custom storage
    custom_storage=surreal_storage,
)

info = auto.run_vector_retriever(
    contents=["https://example.com/agents-blog-post.html"],
    query="How do autonomous agents coordinate tasks?",
    return_detailed_info=True,
)

print(info)

Resources

• Camel documentation

With a lightweight driver you get one database for document chunks, agent memory and vector search—simplifying deployment and scaling for your Camel-powered applications.