Embedding (Vectors)

Turns text into a fixed-size vector. The foundation for RAG, semantic search and similarity comparisons.

Available models

Model	Dimensions	Languages
veriEmbedding (recommended)	1024	TR + EN + 100+ (multilingual)
text-embedding-3-large	3072	Via OpenAI
text-embedding-3-small	1536	Via OpenAI

curl

curl https://app.qevron.ai/v1/embeddings \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{"model": "veriEmbedding", "input": ["hello world", "merhaba dünya"]}'

Python (openai SDK)

from openai import OpenAI
import numpy as np

client = OpenAI(api_key="sk-...", base_url="https://app.qevron.ai/v1")

resp = client.embeddings.create(
    model="veriEmbedding",
    input=["hello world", "merhaba dünya", "qevron is an AI gateway"],
)
vectors = np.array([item.embedding for item in resp.data])
print(vectors.shape)   # (3, 1024)

Batch

input as an array processes all of them in one pass — 10–100× faster than per-item calls:

texts = ["chunk 1", "chunk 2", ..., "chunk 256"]
resp = client.embeddings.create(model="veriEmbedding", input=texts)

Batching pattern

When indexing, send 50–100 chunks per request. One huge request (1000+) optimises throughput but risks timeouts.

Cosine similarity

from numpy import dot
from numpy.linalg import norm

def cosine(a, b):
    return dot(a, b) / (norm(a) * norm(b))

similarity = cosine(vectors[0], vectors[1])   # 0.85+ ⇒ very similar

RAG pipeline (typical)

1. Chunk documents (~500 tokens each)
2. Embed each chunk → veriEmbedding
3. Store (chunk_text, vector) in a vector DB (pgvector, qdrant, ...)
4. On user query → embed query → search vector DB for top-K similar chunks
5. (optional) re-rank with verirerag
6. Pass selected chunks as context to verinova-large for the final answer

End-to-end: RAG walkthrough.

Troubleshooting

• Vectors look identical → input might be empty or too short. Trim whitespace, supply ≥ 5–10 chars.
• Too slow → batch instead of single calls.
• Dimension mismatch → vector DB schema should match the model (1024 for veriEmbedding). Switching from OpenAI (1536/3072) requires a full re-index.

Next: Rerank · RAG walkthrough