The web search API to find, read and cite pages at scale.
It's also self-hostable.
The infrastructure layer that helps AI find, read and cite the live web. Numbers below come from this deployment's index.
One HTTP endpoint per job, JSON in and JSON out, and a dependency-free Python client.
One prompt. Paste this into Claude Code, Cursor or any agent that can read URLs and it wires Nevabase up for you.
Add web search to this project using Nevabase.
Read the API reference at /docs and the
OpenAPI spec at /api/openapi.
Use POST /search with contents.highlights for retrieval,
POST /contents to read pages, and POST /answer when the
user wants a cited answer. Send the key in the x-api-key
header and read it from the NEVABASE_API_KEY env var.from nevabase import Nevabase
nv = Nevabase("nv-...")
r = nv.search(
"startups building AI agents for lawyers",
type="neural", num_results=5,
contents={"highlights": True},
)
for hit in r["results"]:
print(hit["title"], hit["url"])
print(" ", hit["highlights"][0])const r = await fetch("/search", {
method: "POST",
headers: { "x-api-key": "nv-...", "Content-Type": "application/json" },
body: JSON.stringify({
query: "startups building AI agents for lawyers",
type: "neural",
numResults: 5,
contents: { highlights: true },
}),
});
const { results } = await r.json();curl -X POST /search \
-H "x-api-key: nv-..." \
-H "Content-Type: application/json" \
-d '{
"query": "startups building AI agents for lawyers",
"type": "neural",
"numResults": 5,
"contents": { "highlights": true }
}'Agent onboarding. Are you an AI agent? Fetch the OpenAPI spec to discover every endpoint and start building.
curl -s /api/openapicostDollars for meteringSearch infrastructure you can read, run and point at your own corners of the web.
Hybrid ranking. Neural similarity fused with BM25, so “founders on why they shut down” and exact product names both rank right.
See the search options ›Illustrative: hybrid covers both query styles; the others each miss one.
Measured live. The searches below run against this deployment when the page loads, from the browser, through the public API.
Try the playground ›| Query | Search |
|---|---|
| Running… |
Only the passages that matter. Send the model a scored highlight instead of a whole page. Summaries and full text are one flag away.
Contents options ›Code you can run. A single Python service with SQLite storage and CPU embeddings. One command, or the Dockerfile.
Index management API ›./run.sh
# → http://localhost:8000
docker compose up --buildCrawling, cleaning, embedding, ranking and metering are built in, so you only write the prompt.
Search by meaning. Describe the page you want the way you'd describe it to a friend. Filters for domains, dates, category and must-include text.
Clean page text. Boilerplate-free text, query-focused highlights and summaries for any URL or result ID.
Find similar pages. Give us a link and get semantically similar pages: competitor discovery, related reading, dataset expansion.
Cited answers. Search plus synthesis with inline citations to the pages the answer came from. Streams over SSE.
Knows what you meant. Short keyword queries are rewritten into the natural-language form the neural index was built for.
Never a stale miss. Unseen URLs are fetched, cleaned and embedded on demand, then kept for everyone's next query.
Let an agent issue dozens of neural queries, read highlights, and follow up with find-similar.
02Ground a chatbot in fresh pages with cited answers instead of guesses.
03Expose search, read and similar as tools. Responses are small, structured and citation-ready.
04Start from one company's site and map the rest of the market by similarity.
05Find companies matching a description, then pull clean text from each to fill in the record.
06Point the crawler at your own docs or domains and search them with the same API.
Controls mapped to the trust criteria for security, availability and confidentiality. The audit report will be available under NDA once issued.
EU data-subject rights honoured: access, export and deletion of account data on request, with a data processing agreement available.
California consumer privacy rights, including access and deletion requests, handled through the same account-data process.
Safeguards for protected health information via self-hosted deployment inside your own environment. BAA discussions for enterprise plans.
Queries and results are never written to disk or trained on. The usage log keeps only endpoint, timestamp, result count, latency and cost.
DPA documentation, custom data-security arrangements, and fully self-hosted deployment so data never leaves your infrastructure.
url, title, publishedDate, author, score, and optional text, highlights and summary. Every response includes costDollars. See the API reference./contents and /findSimilar live-crawl unknown URLs by default, clean them, embed them and keep them for the next query./answer mode is enabled, the query and retrieved passages are sent to the configured model provider for that request only.Free tier included. No credit card. Self-host any time.