Open in the interactive docs · markdown for agents: https://neuraldeep.tech/docs/vector.md

Vectorization · Embeddings

OpenAI-compatible /v1/embeddings. Accepts a string or an array of strings. The vector size depends on the model: 1024 for e5-large and bge-m3, larger for the others; the exact number is the length of embedding in the response.

curl https://api.neuraldeep.tech/v1/embeddings \
  -H "Authorization: Bearer $YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"e5-large","input":"hello world"}'

curl https://api.neuraldeep.tech/v1/embeddings \
  -H "Authorization: Bearer $YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"bge-m3","input":["text 1","text 2","text 3"]}'

from openai import OpenAI
client = OpenAI(api_key="$YOUR_KEY", base_url="https://api.neuraldeep.tech/v1")
texts = ["first document", "second", "third"]
r = client.embeddings.create(model="bge-m3", input=texts)
vectors = [e.embedding for e in r.data]  # list[list[float]], dim=1024 for bge-m3

For RAG — e5-large for queries/docs with the "query: " / "passage: " prefix. For accuracy — bge-m3 (longer context).

Some models require prefixes. The server does not add them, the client writes them at the start of each text. Without a prefix you still get vectors, but retrieval on them is several times worse.

model	query	document
`frida`	`search_query: …`	`search_document: …`
`e5-large`	`query: …`	`passage: …`
`qwen3-embedding-4b`, `qwen3-embedding-8b`	`Instruct: <task>\nQuery: …`	no prefix

frida has other modes too: paraphrase: for similarity and dedup, categorize: for classification, categorize_sentiment: and categorize_topic: for sentiment and topics.

curl https://api.neuraldeep.tech/v1/embeddings \
  -H "Authorization: Bearer $YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"frida","input":["search_query: how do I get a refund","search_document: Refunds are requested in your account within 14 days"]}'

Input limits.

All text in one request, summed over every input item, up to 105,600 characters (about 32,000 tokens). Beyond that: 400 with the X-Window: vector-payload header; split it into several requests. There is no separate cap on the number of items in the array.
Keep each text within the model window: frida 512 tokens, giga-embeddings 4096, bge-m3 and qwen3-embedding-8b 8192, jina-embeddings-v4 and qwen3-embedding-4b 32,768.
Embeddings and rerank have their own concurrency, separate from chat. Beyond it: 429 with X-Window: vector-inflight and Retry-After. Quotas are separate too; the remainder is in the vector block of GET /v1/limits.

Vectorization · Embeddings

Reference sections (markdown for agents)