With Conare, the same AI agent gets nearly 9 in 10 contract questions right.

Conare is the search an AI agent uses to find the right document before it answers. We gave one agent questions about real contracts and swapped only the search: Conare, OpenAI file search, Cohere or Voyage.

Contract questions answered right

776 questions, the same AI agent, each system's search

  1. Conare89.4%
  2. OpenAI74.5%
  3. Cohere70.2%
  4. Voyage66.9%
One of the 776 contract questions, run with each search system.
Legal agentLegalBench-RAG

Consider Excelerate's Non-Disclosure Agreement; Does the document mention that some obligations of the Agreement may survive the termination of the Agreement?

Each system's first result, out of 714 contracts

  1. ConareExcelerate NDAAnswered right after 1 searchParagraph 10…other Party; however, the obligations of a receiving Party pursuant to Paragraph 4 shall remain in effect for the term …
  2. OpenAIDoiT NDACould not find it after 3 searchesWrong contract
  3. CohereFlexion Therapeutics / Pacira merger agreementCould not find it after 3 searchesWrong contract
  4. VoyageAdamas / Supernus merger agreementCould not find it after 3 searchesWrong contract

Conare is 1st on all 6 benchmarks.

Benchmarks down the side, companies across. Bold is the best score in each row.

BenchmarkConareOpenAINVIDIACohereQwenVoyageZeroEntropyJina
Contract questions1LegalBench-RAG, agent accuracy89.4%1st of 474.5%file search—70.2%embed-v5 + rerank-v5.0-pro—66.9%voyage-4-large + rerank-3——
Contract search2LegalBench-RAG, Recall@585.4%1st of 947.8%file search44.5%Nemotron-3.5-Embed-8B49.4%embed-v5 + rerank-v5.0-pro42.5%Qwen3-Embedding-8B63.7%voyage-context-341.6%zembed-139.9%embeddings-v5-text-small
Charts and tables6ViDoRe v3, RCP-nDCG95.01st of 1777.4embedding-3-large90.2Nemotron-3.5-Rerank-8B92.2rerank-v5.0-pro86.2Qwen3-Reranker-8B90.4rerank-2.591.4zerank-271.6reranker-v3
Reasoning-heavy search5BRIGHT, RCP-nDCG89.21st of 1753.8embedding-3-large60.5Nemotron-3.5-Rerank-8B82.2rerank-v5.0-pro68.9Qwen3-Reranker-4B72.7rerank-2.582.6zerank-254.5reranker-v3
Web search mix8NanoBEIR, RCP-nDCG90.31st of 1779.1embedding-3-large86.1Nemotron-3.5-Rerank-8B87.3rerank-v5.0-pro82.5Qwen3-Reranker-4B85.2rerank-2.587.2zerank-269.2reranker-v3
Agent memory7LMEB, nDCG@1075.01st of 10864.0embedding-3-large72.5Nemotron-3.5-Embed-8B—64.9Qwen3-Embedding-8B—69.1zembed-162.6embeddings-v5-text-small

Conare through its live API. Legal rows: our runs. RCP-nDCG rows: Cohere’s open release of AI-judged scores; OpenAI’s and NVIDIA’s are our runs.9 Data from 13 October to 10 November 2026.

conare.ai/benchmarks

Conare is 1st of 20 in research retrieval (arXiv).

Find one specific paper from a vague description, a detail in its body or a half-remembered title. The score is the share of questions where the right paper is in the top 10 results.

NEEDLE Scholar (arXiv)

Conare 96.2%, 15.2 points ahead of Keenable at 81.0%.

  1. 1Conare96.2%
  2. 2Keenable81.0%
  3. 3context.dev65.7%
  4. 4Parallel61.2%
  5. 5Perplexity60.9%
  6. 6Kagi59.3%
  7. 7Brave58.8%
  8. 8Exa58.0%
  9. 9Jina54.9%
  10. 10Firecrawl54.7%
  11. 11Claude search53.9%
  12. 12Google (Serper)53.5%
  13. 13Linkup52.7%
  14. 14You.com52.3%
  15. 15TinyFish45.8%
  16. 16Bing40.5%
  17. 17Octen36.6%
  18. 18Tavily34.0%
  19. 19DeepSeek27.9%
  20. 20ChatGPT search23.6%

All 20 companies on the board are shown.

NEEDLE is Keenable’s open, live search benchmark. Every company it published for 20 to 25 October 2026, best mode each, on the same 1,960 questions. Conare searches its arXiv index; the other engines search the web.

conare.ai/benchmarks

The full results.

Contract questions

LegalBench-RAG, answered right. An AI agent answers questions about real contracts.

Conare #1 of 4 systems
  1. Conare89.4%
  2. OpenAI74.5%
  3. Cohere70.2%
  4. Voyage66.9%

Share of questions the AI agent answered right

We ran all four systems, with the same agent and AI model for each. Data as of 9 November 2026.1

conare.ai/benchmarks

Conare's first result came from the right contract 92.0% of the time.

Each question is about one contract out of 714. When the first search brings back another contract, the agent starts from the wrong document. OpenAI's first result came from another contract on 49.9% of the questions.

  1. Conare92.0%
  2. Cohere65.7%
  3. Voyage60.8%
  4. OpenAI50.1%

Share of questions where the first search's top result came from the right contract

The right contract was in the top five results for most questions: Conare 98.2%, Cohere 95.7%, OpenAI 94.8%, Voyage 90.2%. What differs is the first result.

From the raw answers: for each question and system, whether the first search's top result came from the contract the question names. Same agent and AI model for every system.1,10 Data as of 9 November 2026.

conare.ai/benchmarks

85 questions only Conare got right. On 78 of them, the other three said “could not find it”.

Same agent, same AI model, same 714 contracts: only the search changed. The other three searches mostly brought back the wrong contract.

ResultConareOpenAICohereVoyage
First result from the right contractOn the 85 questions only Conare got right
First result from the right contractOn the 85 questions only Conare got right77 of 8512 of 8515 of 859 of 85

Merger agreement

Medallia merger: what do shareholders get?

AgentSame AI model

Consider the Acquisition Agreement between Parent "Project Metal Parent, LLC" and Target "Medallia, Inc."; What is the Type of Consideration

Merger agreement

714 contracts

  1. Conare: first result fromMedallia / Thoma Bravo merger agreement, the right contract. Answered right after 1 search
  2. OpenAI: first result fromEaton Vance / Morgan Stanley merger agreement, another contract. Could not find it after 3 searches
  3. Cohere: first result fromCooper Tire / Goodyear merger agreement, another contract. Could not find it after 3 searches
  4. Voyage: first result fromChange Healthcare / UnitedHealth merger agreement, another contract. Could not find it after 3 searches

Conare

Answered right after 1 search

Right contract: Medallia / Thoma Bravo merger agreement. The clause Conare's agent quoted, Section 2.7(a)(iii)

2.7 Effect on Capital Stock. … (iii) each share of Company Common Stock that is issued and outstanding as of immediately prior to the Effective Time (other than Owned Company Shares and Dissenting Company Shares) will be cancelled and extinguished and automatically converted into the right to receive cash in an amount equal to $34.00, without interest thereon (the “Per Share Price”), in accordance with the provisions of Section 2.9 …

Underlined: the words that answer the question

Agent's answer: The consideration is all cash. Under Section 2.7(a)(iii), each share of Company Common Stock is “automatically converted into the right to receive cash in an amount equal to $34.00, without interest thereon”.

Could not find it after 3 searches

First result: Eaton Vance / Morgan Stanley merger agreement, another contract

Without limiting the generality of the foregoing, and subject thereto, at the Second Merger Effective Time, all of the property, rights, privileges, immunities, powers and franchises of the Surviving Corporation and Merger Sub 2 shall vest in the Surviving Company, and all claims, obligations, debts, liabilities and duties of the Surviving Corporation and Merger Sub 2 shall become the claims, obligations, debts, liabilities and duties of the Surviving Company. …

Agent's answer: The passages I retrieved do not state the type of consideration.

Real questions and contracts from the legal test. Conare's panel shows the clause its agent quoted; the others show their first result as returned, where “…” marks a cut in a 500-character excerpt. The map of contracts is a drawing; the documents and passages are real.1

conare.ai/benchmarks

More right, in less time: 10.6 seconds a question.

The same agent and AI model on the same questions. With Conare's search it needed fewer searches and read fewer tokens, and it got more answers right.

Point at a system to read its values off both axes. The vertical axis starts at 60% and the horizontal at 9 seconds.

Averages per question, from the raw answers: seconds logged, searches made, input tokens the AI model read.1,11 Data as of 9 November 2026.

conare.ai/benchmarks

On merger agreements, Conare gets 186 of 194 right and OpenAI 103.

The legal test by document type. Each square is one question, in the same order for every system. The gap is widest on merger agreements; on privacy policies all four are close.

  • Conare186 of 194 right

  • OpenAI103 of 194 right

  • Cohere91 of 194 right

  • Voyage74 of 194 right

SystemMerger agreementsNDAsCommercial contractsPrivacy policies
Conare89.4% right
186 of 194
154 of 194
178 of 194
176 of 194
OpenAI74.5% right
103 of 194
140 of 194
165 of 194
170 of 194
Cohere70.2% right
91 of 194
124 of 194
156 of 194
174 of 194
Voyage66.9% right
74 of 194
112 of 194
160 of 194
173 of 194

  • Answered right, Conare
  • Answered right, others
  • Incomplete or wrong
  • Could not find it
  • One of the Examples

LegalBench-RAG, 194 questions per contract type.1 Graded blind against the passages lawyers marked. Data as of 9 November 2026.

conare.ai/benchmarks

How every benchmark was run.

Three kinds of test. We always ran Conare ourselves, through its live API on default settings; who ran the other systems differs, and each row says so.

  1. 01Our legal testWe ran all four search systems on the same contract questions with one AI agent and one AI model. Only the search changed, and every answer was graded blind against the passage lawyers marked.
  2. 02NEEDLE, the live search benchmarkKeenable’s open benchmark. We ran Conare on its questions with its own scorer. Every other company’s row is as NEEDLE published it, best mode each.
  3. 03Model benchmarksBRIGHT, ViDoRe, NanoBEIR and LMEB. We ran Conare on each benchmark’s own questions and candidates. The other models’ rows are as each release published them.

The systems on the legal test

  • ConareConare live API, default settings
  • OpenAIOpenAI file search, default settings
  • CohereCohere embed-v5 with rerank-v5.0-pro
  • VoyageVoyage voyage-4-large with rerank-3

Notes

  1. 3

    NEEDLE is Keenable's open, live search benchmark. Every company it published for 20 to 25 October 2026, best mode each, with failed searches left out as NEEDLE scores them now. Conare was read live on 9 November 2026 on the same 1,960 questions. Conare searches its arXiv index; the other engines search the web. NEEDLE

  2. 5

    1,366 questions, all 12 BRIGHT tasks. Every reranker orders the same candidates. RCP labels and the other rows: Cohere, open release (CC BY 4.0); OpenAI and NVIDIA: our runs, 12 November 2026. On the original human labels Conare is also first, 47.5 to zerank-1's 38.1. Data as of 21 October 2026. BRIGHT

  3. 6

    2,419 questions over 8 document sets. Every reranker orders the same 150 candidate pages. RCP labels and the other rows: Cohere, open release (CC BY 4.0); OpenAI and NVIDIA: our runs, 12 November 2026. On the original human labels Conare is also first, 70.2 to 68.6. Data as of 13 October 2026. ViDoRe

  4. 7

    13 LMEB tasks with a commercial-use license. Other rows as published on the LMEB leaderboard, 12 October 2026; OpenAI and ZeroEntropy: our runs, 12 November 2026. Data as of 13 October 2026. LMEB

  5. 8

    12 NanoBEIR datasets (MS MARCO left out). Every reranker orders the same candidates. RCP labels and the other rows: Cohere, open release (CC BY 4.0); OpenAI and NVIDIA: our runs, 12 November 2026. On the original human labels Conare is 10th of 17, with 70.4 to the leader's 75.1 (NVIDIA Nemotron 3.5, our run). Data as of 7 November 2026. NanoBEIR

  6. 9

    RCP-nDCG: how well the top 10 results are ordered, from 0 to 100, graded by an AI judge that reads every candidate. Human-label scores use each benchmark's original labels.

  7. 10

    Right contract: the top result of the agent's first search came from the contract the question names, out of 714. Counted for every question and system in the raw answers. The three Examples are from the 85 questions only Conare got right, one from each of three contract types. Data as of 9 November 2026.

  8. 11

    Averages per question over the whole legal test, from the raw answers. Seconds is the time logged for each question; the file does not split it into search time and AI model time. Searches is how many searches the agent made. Input tokens is what its AI model read, with the same AI model for every system. The file has no prices. Data as of 9 November 2026.

Legal questions and grading: LegalBench-RAG (Pipitone and Alami, 2024). NEEDLE and model rows: each benchmark's own release.

conare.ai/benchmarks

Questions about these results

  • Conare is a retrieval API for AI agents and RAG apps. It stores your data, keeps it in sync and returns the right context to your AI, across 100k to over 100M documents, through the API, the SDK or MCP.

  • One agent and one AI model. The legal test is LegalBench-RAG's mini set: questions about 72 public contracts, searched across a collection of 714. OpenAI file search ran on its default settings; Cohere and Voyage are their embedding and reranking models, as named. Answers were graded blind against the passages lawyers marked.

  • An AI-judged score of how well the top 10 results are ordered, from 0 to 100. The AI judge reads every candidate, and the labels come from Cohere's open release. The notes also give each benchmark's original human-label score.

  • The table above marks every row another system leads, in bold. Head to head on the legal test, OpenAI was right where Conare missed on 24 questions, Cohere on 19 and Voyage on 21.

  • Yes. Every answer from the legal test, for all four systems, is in one file with the grade and the reason for it. Ask us for it when you book a demo.

  • Conare is SOC 2 Type II and GDPR compliant, and we do not train models on your content. The details are at trust.conare.ai.

  • Conare is invite-only for now. Book a demo: we set up your organization with you, and your team then creates API keys in the console.

Run it on your own documents.

Ask anything
⌘K