The problem
Keyword search can miss a useful passage because the question uses different words. A chat model also cannot remember every document in your system between requests. Your agent needs a retrieval step that can find relevant material before it answers.
How NativePort helps
NativePort gives your application access to Jina embeddings, which turn text into numerical representations of its meaning. Store these alongside your documents, then compare a question’s representation to find useful passages for the agent.
Technical implementation
Use Python 3 with requests installed (python -m pip install requests). Set NATIVEPORT_API_KEY in your environment to your NativePort key. Run the snippets on your server, where your key stays private.
This example creates vectors for two passages, then embeds the question with the matching query task.
import os
import requests
BASE = "https://api.nativeport.ai"
HEADERS = {"Authorization": f"Bearer {os.environ['NATIVEPORT_API_KEY']}"}
def embed(texts, task):
response = requests.post(BASE + "/jina/embeddings", headers=HEADERS,
json={"model": "jina-embeddings-v3", "task": task,
"input": texts, "normalized": True}, timeout=90)
response.raise_for_status()
return [item["embedding"] for item in
sorted(response.json()["data"], key=lambda item: item["index"])]
passages = ["Returns are accepted within 30 days.", "Support opens at nine."]
vectors = embed(passages, "retrieval.passage")
query = embed(["When can I return a purchase?"], "retrieval.query")[0]
scores = [sum(a * b for a, b in zip(query, vector)) for vector in vectors]
best = max(range(len(scores)), key=scores.__getitem__)
print(passages[best])For larger collections, store vectors in your own vector index such as pgvector or OpenSearch. Keep the same model and dimensions for stored passages and queries; a model change generally requires re-embedding. The chat endpoint does not generate embeddings. See Jina’s routes.
Tool costs
| Tool used in the example | Price per call |
|---|---|
| Jina — text embeddings | $0.0000211 / 1,000 tokens (usage-based) |
Prices in USD. Usage-based tools have no fixed per-call price. View pricing.
Let your agent build it
You don’t need to write this code yourself. Copy this page’s link and paste it into your agent. Ask it to follow the guide and implement the feature for you.
Keep exploring
Answer questions from supplied evidence
Give a model source passages so it can explain an answer and point back to the evidence.
View guide
Choose the model your agent needs
Use one chat interface to call available models across providers, including open models.
View guide
Classify text using your own review criteria
Ask a language model to label sentiment and identify content that needs review.
View guide
Explore all use cases
Find the next tool your agent needs, from web access to voice and document processing.
All use cases