Llama Index

Llama Index is a framework for building RAG pipelines. It provides a flexible and powerful way to build and deploy RAG systems.

Note

Llama Index doesn’t (yet!) ship an off-the-shelf SurrealDB adapter, so in this guide we’ll implement a 20-line SurrealVectorStore that satisfies Llama Index’s minimal vector-store interface. You can drop it into any RAG pipeline just like you would with Pinecone, Qdrant, Chroma, etc.

TL;DR

Install surrealdb + llama-index.
Create a table with an HNSW index.
Drop-in the 20-line SurrealVectorStore.
Use VectorStoreIndex.from_documents() and you’re done.

Install the libraries

pip install surrealdb llama-index numpy

Connect to SurrealDB

from surrealdb import Surreal
import asyncio, os

# — connection details —
DB_URL   = os.getenv("SDB_URL", "http://localhost:8000/rpc")
DB_USER  = os.getenv("SDB_USER", "root")
DB_PASS  = os.getenv("SDB_PASS", "secret")
NS, DB   = "demo", "demo"
TABLE    = "llama_chunks"

sdb = Surreal(DB_URL)

async def init():
    await sdb.signin({"user": DB_USER, "pass": DB_PASS})
    await sdb.use(NS, DB)

asyncio.run(init())

Create the table and HNSW index

# SurrealDB ≥ v1.5 exposes an in-memory HNSW ANN index
schema = """
DEFINE TABLE $tb SCHEMALESS PERMISSIONS NONE;
DEFINE FIELD text      ON $tb TYPE string;
DEFINE FIELD embedding ON $tb TYPE array;

DEFINE INDEX idx_hnsw ON $tb
  FIELDS embedding
  HNSW DIMENSION 768
  DIST   COSINE;
"""
asyncio.run(sdb.query(schema, {"tb": TABLE}))

SurrealQL’s DEFINE INDEX … HNSW activates a high-speed, cosine-distance ANN index.

A tiny SurrealDB adapter for Llama Index

import numpy as np
from typing import Any, List, Sequence
from llama_index.core.vector_stores.types import (
    BasePydanticVectorStore,
    VectorStoreQuery,
    VectorStoreQueryResult,
)
from llama_index.core.schema import NodeWithEmbedding

class SurrealVectorStore(BasePydanticVectorStore):
    """Minimal Llama Index adapter using SurrealDB’s <|K,EF|> operator."""

    def __init__(self, table: str, conn: Surreal):
        self.table, self.conn = table, conn

    # ---- required APIs ----------------------------------------------------
    def add(self, nodes: Sequence[NodeWithEmbedding], **_) -> List[str]:
        rows = [
            {
                "id": f"{self.table}:{i}",
                "text": n.get_content(metadata_mode="all"),
                "embedding": n.embedding,
            }
            for i, n in enumerate(nodes)
        ]
        asyncio.run(self.conn.query(f"INSERT INTO {self.table} $data", {"data": rows}))
        return [r["id"] for r in rows]

    def delete(self, doc_id: str, **_) -> None:  # optional but easy
        asyncio.run(self.conn.delete(doc_id))

    def query(
        self, query: VectorStoreQuery, **_
    ) -> VectorStoreQueryResult:  # semantic search
        q_vec = query.query_embedding
        surql = """
        LET $q := $vec;
        SELECT id, text, vector::distance::knn() AS score
        FROM {self.table}
        WHERE embedding <|{query.similarity_top_k},64|> $q
        ORDER BY score;
        """
        res = asyncio.run(self.conn.query(surql, {"vec": q_vec}))[0].result
        return VectorStoreQueryResult(
            nodes=[],  # Llama Index will fetch raw docs separately
            ids=[r["id"] for r in res],
            similarities=[r["score"] for r in res],
        )

The SurrealVectorStore class is a minimal adapter cribbed from the official build a vector store from scratch example.

Index documents with Llama Index

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader

# 5-1  load + chunk a handful of docs (replace with your own)
docs = SimpleDirectoryReader("./my_pdfs").load_data()

# 5-2  build the index – Llama Index will call SurrealVectorStore.add()
surreal_store = SurrealVectorStore(table=TABLE, conn=sdb)
index = VectorStoreIndex.from_documents(
    docs,
    vector_store=surreal_store,
    show_progress=True,
)

Query

qe   = index.as_query_engine(similarity_top_k=4)
resp = qe.query("Which document talks about vector search in SurrealDB?")
print(resp)

The query flow is:

User → Llama Index → SurrealVectorStore.query()
          ↘︎ top-K doc IDs & scores
                  ↘︎ fetch full text → synthesize answer

What about metadata filters, deletes, hybrid search?

SurrealDB supports metadata-rich JSON payloads, additional filter clauses, and full-text search; extend SurrealVectorStore.query() to:

add WHERE predicates before the <|K,EF|> operator,
combine with vector::distance::knn() for re-ranking,
or fall back to vector::distance::cosine() for exact search.

All other Llama Index abstractions (retrievers, query engines, agents) work unchanged, because they talk to the vector store through the same tiny interface you just implemented.

Enjoy Llama Index + SurrealDB!