
All-in-one API for search, RAG, & recommendations
Turn this teardown into a decision-ready prompt for ChatGPT, Claude, or your agent.
If you only have a few minutes to spare, here’s what investors, operators, and founders should know about Trieve (W24).
Trieve was a Rust-based search, retrieval, recommendation, and RAG platform founded in 2023 by Nick Khami and Denzell Ford. The company raised $3.5 million and sold to Mintlify in a completed transaction announced July 24, 2025.[1][2]
This was an acquisition success, not a failure story. Mintlify adopted Trieve before the founders entered YC, relied on it for documentation search, and later bought the team and retrieval technology as search became central to its product. Survival split after the deal: Trieve's code moved from the Business Source License to MIT and remains public, while Trieve Cloud shut down on November 1, 2025. Repository availability is proven; continuing feature investment is not.
Nick Khami and Denzell Ford met as freshman research partners at the University of Texas at Austin while building a machine-learning model to predict oil locations.[3] They later built semantic retrieval for a knowledge-management product meant to help business teams make stronger arguments. At acquisition, Khami described the first application more narrowly as relevancy-optimized retrieval for litigation attorneys.[4] The descriptions are compatible, but the public record does not establish the earlier product's name, customers, or full feature set.
The pivot came from buyer pull. Prospects did not primarily want the argument application; they wanted the semantic-search infrastructure underneath it.[1] Trieve became a general retrieval platform spanning search, recommendations, and RAG. The founders' aspiration was concise: “Power every retrieval experience on the internet.”[1]

Trieve joined the FUTO Fellows Program in 2023 before YC's Winter 2024 batch.[5] Mintlify found the company before YC after an early retrieval demonstration showed strong quality and performance, according to the buyer.[2]
The canonical format asks for two exact founder quotes. The corpus preserves the ambition line above, but not a second verified verbatim passage suitable for reproduction. No speech is reconstructed.
Trieve wrapped a vector database in the application machinery required for production retrieval. Its backend was primarily Rust and used Qdrant for indexing and vector search. Around that core, Trieve built ingestion, multitenancy, chunk grouping, filtering, relevance weights, recency biasing, recommendations, and RAG endpoints.[6]

The retrieval path combined dense embeddings, SPLADE sparse vectors, and cross-encoder reranking. Users could replace Trieve's embedding, SPLADE, reranking, or language models with their own.[10] This made relevance tunable rather than fixed to one provider. The product also supported typo-tolerant search, sub-sentence highlighting, grouping, filters, managed-context and user-selected-context RAG, and recommendations.
Read the complete post-mortem, the rebuild playbook, and the exact reasons Trieve is still worth studying now.