Weaviate

import os

import weaviate
import weaviate.classes.config as wc
from openai import OpenAI

# --- генерация: SpeShu.AI ---
speshu = OpenAI(
    base_url="https://speshu.ai/api/v1",
    api_key=os.environ["SPESHU_API_KEY"],
)

# --- векторы: отдельный провайдер эмбеддингов, НЕ SpeShu.AI ---
embeddings_client = OpenAI(
    base_url=os.environ["EMBEDDING_BASE_URL"],
    api_key=os.environ["EMBEDDING_API_KEY"],
)
EMBEDDING_MODEL = os.environ["EMBEDDING_MODEL"]

# --- хранилище векторов ---
weaviate_client = weaviate.connect_to_local()

export SPESHU_MODEL=$(curl -s https://speshu.ai/api/v1/models \
  -H "Authorization: Bearer $SPESHU_API_KEY" \
  | jq -r '.data[0].id')

import os

import requests

BASE = "https://speshu.ai/api/v1"

response = requests.get(
    f"{BASE}/models",
    headers={"Authorization": f"Bearer {os.environ['SPESHU_API_KEY']}"},
    timeout=30,
)
response.raise_for_status()
MODEL = response.json()["data"][0]["id"]

import weaviate.classes.config as wc

COLLECTION = "Article"

if weaviate_client.collections.exists(COLLECTION):
    weaviate_client.collections.delete(COLLECTION)

articles = weaviate_client.collections.create(
    name=COLLECTION,
    properties=[
        wc.Property(name="text", data_type=wc.DataType.TEXT),
        wc.Property(name="source", data_type=wc.DataType.TEXT),
    ],
    vector_config=wc.Configure.Vectors.self_provided(),
)

texts = [
    "SpeShu.AI отдаёт текстовые модели по OpenAI-совместимому API.",
    "Weaviate хранит векторы и properties для семантического поиска.",
]

vectors = embeddings_client.embeddings.create(model=EMBEDDING_MODEL, input=texts).data

for text, item in zip(texts, vectors):
    articles.data.insert(properties={"text": text, "source": "manual"}, vector=item.embedding)

question = "Как хранить документы для RAG?"

query_vector = embeddings_client.embeddings.create(
    model=EMBEDDING_MODEL, input=question
).data[0].embedding

hits = articles.query.near_vector(near_vector=query_vector, limit=3)
context = "\n\n".join(obj.properties["text"] for obj in hits.objects)

completion = speshu.chat.completions.create(
    model=MODEL,
    messages=[
        {"role": "system", "content": "Отвечай только по приведённому контексту."},
        {"role": "user", "content": f"Контекст:\n{context}\n\nВопрос: {question}"},
    ],
)

print(completion.choices[0].message.content)
print("списано:", completion.usage.cost.total_cost, "₽")

# Отбор по свойствам: near_vector не вызывается, эмбеддинги не нужны
recent = articles.query.fetch_objects(
    filters=weaviate.classes.query.Filter.by_property("source").equal("manual"),
    limit=10,
)

context = "\n\n".join(obj.properties["text"] for obj in recent.objects)

import weaviate.classes.config as wc

generated = weaviate_client.collections.create(
    name="ArticleGenerated",
    properties=[wc.Property(name="text", data_type=wc.DataType.TEXT)],
    vector_config=wc.Configure.Vectors.self_provided(),
    generative_config=wc.Configure.Generative.openai(
        generative_config={
            "model": MODEL,
            "base_url": "https://speshu.ai/api/v1",
        },
    ),
)