EN
Back to the archive

The archive · AI & Models · Product decision · 2024–2026

ZeroEntropy bets retrieval, not LLMs, is AI's weak link: EF pre-seed, YC W25, $4.2M

RAG apps fail at retrieval, not generation: Ghita Houir Alami built ZeroEntropy through EF and YC, raising a $4.2M seed for a 'Supabase for search'.

ZeroEntropy

The betThat retrieval, not the LLM, is where AI apps break, so teams will swap stitched vector stacks for one 'Supabase for search' API owning ingestion, ranking and evaluation.Live

What the business is

Developer-first search infrastructure: an API that ingests messy documents, indexes them, reranks results and evaluates quality, so AI agents retrieve accurate context instead of hallucinating.

Starting capitalEntrepreneurs First pre-seed, then a $4.2M seed led by Initialized Capital with Y Combinator, Transpose Platform, 22 Ventures and a16z Scout (announced 2025-07-08); about $4.4M total.

How it started

Ghita Houir Alami, a Moroccan mathematician trained at École Polytechnique and UC Berkeley, kept hitting the retrieval wall while building an AI assistant before ChatGPT existed. CTO Nicholas Pipitone, a competitive programmer and CMU-trained builder, had shipped AI agents across multiple startups. Both saw the same failure — wrong context in, hallucination out — so in 2024 they founded ZeroEntropy, pre-seeded by Entrepreneurs First, and joined Y Combinator's Winter 2025 batch.

What happened

ZeroEntropy shipped zerank-1 rerankers, one fully open-source under Apache 2.0, claiming benchmark gains over Cohere and Salesforce models. In July 2025 it raised a $4.2M seed led by Initialized Capital, with Y Combinator, Transpose Platform, 22 Ventures, a16z Scout and angels from OpenAI, Hugging Face and Front participating. TechCrunch reported that more than 10 early AI-agent companies across healthcare, law, customer support and sales were already using the API, which the founders described as a 'Supabase for search' handling ingestion, indexing, re-ranking and evaluation behind one endpoint.

No ending yet — it is still running.

Background

ZeroEntropy's bet is that retrieval — not the model — is where retrieval-augmented generation breaks. Founder Ghita Houir Alami, a Moroccan mathematician who left home at 17 for École Polytechnique and later studied applied math at UC Berkeley, hit that wall building an AI assistant before ChatGPT existed; CTO Nicholas Pipitone saw the same failure across AI startups. In 2024 they founded ZeroEntropy to fix retrieval once, for everyone.

The path ran through two accelerators: Entrepreneurs First pre-seeded the company, and Y Combinator's Winter 2025 batch took it in. On 2025-07-08 ZeroEntropy announced a $4.2M seed led by Initialized Capital, with Y Combinator, Transpose Platform, 22 Ventures, a16z Scout and angels from OpenAI, Hugging Face and Front. TechCrunch covered the round the next day, reporting that more than 10 early AI-agent companies in healthcare, law, support and sales already used the API.

The product is a 'Supabase for search': one API that handles document ingestion, indexing, hybrid retrieval, re-ranking and evaluation, replacing the default stack of off-the-shelf embeddings, vector search and generic rerankers. Its proprietary reranker, zerank-1, was claimed to outperform Cohere and Salesforce models on public and private retrieval benchmarks. As of 2026-09-02 the company is still live and the bet that developers will pay for a dedicated retrieval layer remains in progress.

What has to be true

  • A legible bet with a dated arc: founded 2024, EF pre-seed, YC W25, then a $4.2M seed on 2025-07-08 documented by TechCrunch, Initialized Capital and Entrepreneurs First.
  • The contrarian claim is concrete: RAG fails at retrieval, not generation, so infrastructure rather than another chatbot is the wedge.
  • Traction is specific: TechCrunch coverage, a Product Hunt launch, an EF portfolio highlight, and 10+ AI-agent startups using the API by July 2025.
  • The founder story — Morocco to École Polytechnique to UC Berkeley to YC — is distinctive and verified by TechCrunch.

What can be applied

When every team reinvents the same broken pipeline, owning that layer beats building another app; but infra startups live on benchmarks and trust, and every release must re-earn developers.

Aftermath

As of 2026-09-02 ZeroEntropy is still running, selling retrieval infrastructure to AI-agent builders; no shutdown, exit or later funding round appears in the verified sources. The company has continued shipping models under its zerank and zembed lines, but the central bet — that teams will replace hand-stitched retrieval stacks with one API — is still being tested in the market.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases