Embeddings

Endpoint

http
POST https://api.aicodedog.com/v1/embeddings

Popular models: text-embedding-3-large (3072 dim, top quality), text-embedding-3-small (1536 dim, best value), bge-m3 (strong on both Chinese and English).

Minimal request

curl
curl https://api.aicodedog.com/v1/embeddings \
  -H "Authorization: Bearer $AICODEDOG_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "text-embedding-3-small",
    "input": "An API relay lets you call many models with a single key"
  }'

python
from openai import OpenAI
client = OpenAI(api_key="<KEY>", base_url="https://api.aicodedog.com/v1")

resp = client.embeddings.create(
    model="text-embedding-3-small",
    input=["first paragraph", "second paragraph", "third paragraph"],  # batches also work
)
vectors = [d.embedding for d in resp.data]
print(len(vectors[0]))  # 1536

Parameters

• model (required) — model slug.
• input (required) — a string or an array of strings. Batches up to 2048 items per request are most efficient.
• dimensions — truncation length (v3 series only). Setting 512 drops 1536-dim vectors to 512 dims, storage cost goes down to 1/3.
• encoding_format — float (default) or base64 (saves bandwidth).

Cosine similarity

Once you have two vectors, cosine similarity is the standard "how close are they?" metric:

python
import numpy as np

def cosine(a, b):
    a, b = np.array(a), np.array(b)
    return float(a @ b / (np.linalg.norm(a) * np.linalg.norm(b)))

# 1.0 = identical; 0.0 = unrelated; -1.0 = opposite
print(cosine(vectors[0], vectors[1]))

Typical use: RAG retrieval

1. Chunk every document, embed each chunk, store the vectors in a vector DB (Pinecone / Weaviate / pgvector / Qdrant).
2. When a user asks a question, embed the question too and pull the top-k most similar chunks from the DB.
3. Splice those chunks into the system prompt and let chat completions answer based on "the retrieved material".

That's what people call RAG — Retrieval Augmented Generation.