External Collection — BYO embeddings

from qdrant_client import QdrantClient
 
qd = QdrantClient(url="https://qdrant.uk-lon-1.dodil.io", api_key="kb-prod")

from pinecone import Pinecone
 
# The Pinecone adapter takes the TABLE NAME from an X-Dodil-Index header,
# not from the index name in the SDK — without it every call targets a
# table literally named "vectors".
pc = Pinecone(
    api_key="kb-prod",                       # = the db id
    host="https://pinecone.uk-lon-1.dodil.io",
    additional_headers={
        "X-Dodil-Index": "ada",              # the table to read/write
        "X-Dodil-Metric": "cosine",          # cosine | euclidean | dotproduct
    },
)
index = pc.Index(host="https://pinecone.uk-lon-1.dodil.io")

from qdrant_client.models import VectorParams, Distance
 
qd.create_collection(
    collection_name="ada",
    vectors_config=VectorParams(size=1536, distance=Distance.COSINE),
)

CREATE TABLE IF NOT EXISTS ada (
  id     VARCHAR PRIMARY KEY,
  vector VECTOR(1536)
);

import openai
 
docs = [
    {"id": "doc-1", "text": "Multi-head attention lets the model jointly attend to information from different representation subspaces."},
    {"id": "doc-2", "text": "BERT uses bidirectional self-attention over masked tokens."},
    {"id": "doc-3", "text": "GPT uses causal self-attention with left-to-right context."},
]
 
client = openai.OpenAI()
resp = client.embeddings.create(
    model="text-embedding-ada-002",
    input=[d["text"] for d in docs],
)
 
qd.upsert(
    collection_name="ada",
    points=[
        {
            "id": doc["id"],
            "vector": emb.embedding,                       # 1536 floats
            "payload": {"text": doc["text"], "source": "papers/attention.pdf"},
        }
        for doc, emb in zip(docs, resp.data)
    ],
)

index.upsert(vectors=[
    {"id": doc["id"], "values": emb.embedding, "metadata": {"text": doc["text"]}}
    for doc, emb in zip(docs, resp.data)
])

curl -sS -X POST "https://table.uk-lon-1.dodil.io/v1/vector/ada/upsert" \
  -H "X-DB-Id: kb-prod" \
  -H "Content-Type: application/json" \
  -d '{
    "column": "vector",
    "points": [
      {"id": "doc-1", "values": [0.12, -0.04, "…"]}
    ]
  }'

qresp = client.embeddings.create(model="text-embedding-ada-002", input="multi-head attention")
 
hits = qd.search(
    collection_name="ada",
    query_vector=qresp.data[0].embedding,
    limit=5,
    with_payload=True,
)
for h in hits:
    print(h.id, h.score, h.payload.get("text"))

-- <=> cosine · <-> L2 · <#> inner product
SELECT id
FROM ada
ORDER BY vector <=> '[0.12, -0.04, ...]'
LIMIT 5;

SELECT a.id, p.payload
FROM ada a
JOIN ada_payload p ON p.id = a.id
ORDER BY a.vector <=> '[0.12, -0.04, ...]'
LIMIT 5;

dodil data vsearch -b kb-prod -t ada --column vector \
  --vector "0.12,-0.04,…" --metric cosine --top-k 5

from qdrant_client.models import Filter, FieldCondition, MatchValue
 
hits = qd.search(
    collection_name="ada",
    query_vector=qvec,
    query_filter=Filter(must=[FieldCondition(key="source", match=MatchValue(value="papers/attention.pdf"))]),
    limit=5,
)

index.query(
    vector=qvec,
    top_k=5,
    filter={"source": {"$eq": "papers/attention.pdf"}, "year": {"$gte": 2017}},
    include_metadata=True,
)

qd.delete(collection_name="ada", points_selector={"points": ["doc-1", "doc-2"]})
 
# An EMPTY filter is Qdrant's match-all — this clears the whole collection.
qd.delete(collection_name="ada", points_selector={"filter": {}})
 
# Drop the table (and its companions)
qd.delete_collection("ada")

# By id
curl -sS -X POST "https://table.uk-lon-1.dodil.io/v1/vector/ada/delete" \
  -H "X-DB-Id: kb-prod" -H "Content-Type: application/json" \
  -d '{"ids": ["doc-1", "doc-2"]}'
 
# Everything
curl -sS -X POST "https://table.uk-lon-1.dodil.io/v1/vector/ada/delete" \
  -H "X-DB-Id: kb-prod" -H "Content-Type: application/json" \
  -d '{"all": true}'