Anthropic (Claude) RAG Starter

Although Anthropic doesn’t ship an embedding model of its own (it recommends Voyage AI instead), the recipe below uses Voyage AI to embed text, stores the vectors in SurrealDB, and calls Anthropic Claude for generation.

Install the SDKs

pip install surrealdb voyageai anthropic

Connect to SurrealDB

from surrealdb import Surreal
import voyageai, anthropic, asyncio, os

DB_URI     = "http://localhost:8000/rpc"
DB_USER    = "root"
DB_PASS    = "root"
NAMESPACE  = "demo"
DATABASE   = "demo"
TABLE_NAME = "rag_docs"

VOYAGE_API_KEY   = os.environ["VOYAGE_API_KEY"]
ANTHROPIC_API_KEY = os.environ["ANTHROPIC_API_KEY"]

db  = Surreal(DB_URI)
vo  = voyageai.Client(api_key=VOYAGE_API_KEY)
claude = anthropic.Anthropic(api_key=ANTHROPIC_API_KEY)

async def init():
    await db.signin({"user": DB_USER, "pass": DB_PASS})
    await db.use(NAMESPACE, DATABASE)

asyncio.run(init())

Define table + vector index (raw query with variable binding)

schema = """
DEFINE TABLE $tb SCHEMALESS PERMISSIONS NONE;
DEFINE FIELD embedding      ON $tb TYPE array;
DEFINE FIELD original_text  ON $tb TYPE string;

-- HNSW ANN index (SurrealDB ≥ v1.5):contentReference[oaicite:1]{index=1}
DEFINE INDEX idx_hnsw ON $tb
  FIELDS embedding
  HNSW DIMENSION 1024
  DIST   COSINE;
"""

asyncio.run(db.query(schema, { "tb": TABLE_NAME }))   # <- variable binding pattern:contentReference[oaicite:2]{index=2}

Embed and ingest documents

docs = [
    "SurrealDB is a multi-model database that now has an in-memory HNSW index.",
    "Anthropic’s Claude 3 model family (Haiku, Sonnet, Opus) excels at reasoning tasks."
]

doc_vecs = vo.embed(
    docs,
    model="voyage-3",
    input_type="document"
).embeddings          # vectors are 1024-D by default:contentReference[oaicite:3]{index=3}

async def ingest():
    for i, (text, vec) in enumerate(zip(docs, doc_vecs), start=1):
        await db.create(f"{TABLE_NAME}:{i}", {
            "original_text": text,
            "embedding":     vec,
        })

asyncio.run(ingest())

Search SurrealDB with a query

query = "Does SurrealDB support HNSW for fast ANN search?"
q_vec  = vo.embed([query], model="voyage-3", input_type="query").embeddings[0]

surql = """
LET $q := $vec;
SELECT id, original_text, vector::distance::knn() AS dist
FROM $tb
WHERE embedding <|3,64|> $q          -- top-3, efSearch = 64
ORDER BY dist;
"""

hits = asyncio.run(db.query(surql, { "vec": q_vec, "tb": TABLE_NAME }))[0].result

<|K,EF|> is SurrealQL’s K-NN operator; vector::distance::knn() returns the pre-computed cosine distance.

Ask Claude with retrieved context

context = "\n\n".join([h["original_text"] for h in hits])

prompt = [
    { "role": "user",
      "content": `Answer the question using ONLY the context.

Context:
{context}

Question: {query}
"""}
]

response = claude.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=256,
    messages=prompt,
)

print(response.content[0].text)

Done! You now have a lightweight RAG stack:

1. Voyage AI → high-quality embeddings.
2. SurrealDB → blazing-fast HNSW vector search.
3. Claude (Anthropic) → answer generation over the retrieved snippets.