History

Hypar started as something different.

v0.4 — RAG reference implementation (2025–2026)

The original Hypar was a production-ready Retrieval-Augmented Generation application built with Nuxt 3, pgvector, and the Vercel AI SDK. The goal was to build a RAG system readable enough to learn from — every stage of the pipeline documented, every parameter tunable at runtime.

It was a serious implementation: hybrid search (vector + BM25), HyDE, MMR reranking, durable ingestion with Workflow SDK, multi-user auth with better-auth, source citations, an admin dashboard, and a 10-stage roadmap toward a live product.

That work is preserved in its entirety:

• Branch: archive/rag-v1
• Tag: v0.4-rag

The pivot (August 2026)

Working on the RAG implementation surfaced a deeper question: not how to build AI knowledge tools, but what they should feel like to use. The chat interface for RAG felt like the wrong abstraction — a retrieval system dressed as a conversation.

The question became: what does a knowledge system look like when the AI agent is a collaborator in the thinking process, not a search engine over your documents?

Hypar became the place to explore that question. The RAG implementation was archived, not deleted — it represents real engineering work and a specific answer to a specific problem. That answer was just not the question Hypar wanted to keep asking.

What carried over

The infrastructure built for the RAG implementation — Nuxt 3, Tailwind v4, the visual system, the deployment setup — continues to serve the new project. The code changed; the foundation remained.

What shipped after the pivot (August 2026)

On main: embryo garden and detail, fossil UX with kinds, typed connections and a graph surface, streaming collaborator with auto-engage and reply loop, method stance by lifecycle (define / probe / variety / simplest), pending paths, garden question queue, and Ollama model selector. RAG eval CI, embedding docs, and the old chat/document routes are gone.

Research write-ups: method as process, ai-elements surfaces. Open tensions: Open Questions.