How they use Tavily
Keenable uses Tavily as an evaluated search engine in NEEDLE, its live open-source benchmark for agentic web search. Its announcement names Tavily among the search engines compared using refreshed queries.
NEEDLE covers news, finance, scholarly literature, rare entities and legal searches. It builds queries from RSS feeds, Google Trends, financial and scientific APIs, and public agentic logs. Tasks rerun daily or hourly on new queries. Scheduled runs execute in GitHub Actions.
The evaluation protocol sends requests sequentially without concurrent load. It judges each engine’s returned rankings, titles and snippets without fetching pages or reranking, and applies the same snippet-length cap across engines. The current methodology scores engine-caused failures as zero and excludes failures caused by Keenable’s own account or plan from score means. Reports list both kinds of error.
The repository exposes Tavily search-depth tuning through TAVILY_DEPTH. The dashboard documents Tavily requests with search_depth=basic and max_results=min(k,20). Tavily responses are clipped to a 2,000-character snippet budget before judging and answer matching. Requests have a 30-second timeout and retry up to three times with exponential backoff on transport errors, HTTP 429 and 5xx responses.
One recorded benchmark run reports 20 scored Tavily queries, zero search or judge errors, mean NDCG of 0.3211 and mean latency of 2,072.1 ms. That run records median latency of 1,860.3 ms and p95 latency of 3,141.5 ms.
For the seven-day window ending October 31, 2026, the results dashboard reports Tavily at 41% of the benchmark’s ceiling, with a 95% confidence interval of 40–43%, median search latency of 1.8 seconds and p95 latency of 3.4 seconds.
For the seven-day window ending November 4, 2026, at 04:05 UTC, it reports Tavily at 43% of the ceiling, with a 95% confidence interval of 41–45%, median search latency of 1.6 seconds and p95 latency of 3.3 seconds. The synthetic ceiling pools the best results found across engines and ranks them using the same judge that scores the comparison. This relationship concerns benchmark evaluation.
Sources
7 sourcesWeb Search API Benchmark 2026: Keenable vs Exa and 22 APIs
Listed 31 Oct 2026 · keenableai.github.io · Web page
Keenable’s NEEDLE benchmark evaluates Tavily. Its methodology scores engine-caused search failures as zero and excludes failures caused by Keenable’s own account or plan from score means. For the seven-day window ending November 4, 2026, at 04:05 UTC, it reports Tavily at 43% of the ceiling, with a 95% confidence interval of 41–45%, median latency of 1.6 seconds and p95 latency of 3.3 seconds.
Open original sourceNEEDLE live benchmark results
Listed 25 Oct 2026 · Keenable · Web page
Keenable’s NEEDLE benchmark calls Tavily’s search API with search_depth=basic and max_results=min(k,20). It clips Tavily responses to a 2,000-character snippet budget before judging and answer matching. Tavily requests use a 30-second timeout and retry up to three times with exponential backoff on transport errors, HTTP 429 and 5xx responses.
Open original sourcebench: 2026-10-23T11:05Z · keenable-ai/needle-results at 9ffc438
Listed 28 Oct 2026 · huggingface.co · Web page
Keenable’s recorded NEEDLE benchmark run reports 20 scored Tavily queries with zero search or judge errors, mean NDCG of 0.3211, mean latency of 2,072.1 ms, median latency of 1,860.3 ms and p95 latency of 3,141.5 ms.
Open original sourceNEEDLE: The benchmark your search engine can't memorize - Keenable.ai
Published 27 Aug 2026 · keenable.ai · Web page
Keenable compares Tavily with other search engines in NEEDLE. Its synthetic upper bound pools the best results found across engines and ranks them using the same judge that scores the comparison.
Open original sourceKeenable NEEDLE repository
Listed 25 Oct 2026 · GitHub · Web page
Keenable’s NEEDLE repository documents one request at a time per engine and exposes Tavily search-depth tuning through TAVILY_DEPTH. Its continuous benchmark runs news hourly and the other suites daily.
Open original sourceCompany context
keenable.aiKeenable, operated by Keenable.ai, Inc., builds independent web search and knowledge access infrastructure for AI labs, inference providers and agent developers. Its Search API searches web pages and retrieves page content through developer interfaces including MCP. Keenable SELECT uses Web Query Language to turn information from multiple pages into structured tables with source references. Its Time Machine API, available through early access, searches historical web snapshots. The company offers cloud access and dedicated capacity, including on-premises deployment for larger customers.
Recent updates
8 updatesKeenable opens AI Startups program offering $5,000 in credits
Keenable opened applications for its AI Startups program, offering selected teams $5,000 in credits, onboarding, and launch promotion support. The credits are valid for six months and apply across Keenable products.
KeenableKeenable launches Dynamic Snippets
Dynamic Snippets retrieves relevant passages from across long documents. Keenable reports roughly halving search-result tokens on production workloads, with the system enabled by default in its realtime search API.
KeenableKeenable becomes a web-fetch backend for cognee
Cognee version 1.5 adds Keenable as a backend for retrieving web content into agent memory. The integration returns Markdown from indexed or live pages and provides controls over content selection and freshness.
Keenable