Skip to content
← all systems

eve-sage

Agentic RAG agent that is an expert on the framework it is built with - retrieval engineering and eval rigor as the product.

multi-hop retrieval · evals in CI

role
Architect and sole engineer
stack
eve (Vercel) · TypeScript · AI SDK · RAG
status
oss

eve-sage is an agentic RAG agent that is an expert on the framework it is built with - Vercel's eve. Hard questions get decomposed and re-searched until they are answerable, answers carry inline citations grounded in the eve, AI SDK, and Workflow SDK docs, and a finite, high-quality corpus keeps its eval numbers honest. The eval harness runs in CI, so a retrieval-quality regression fails the build - the retrieval rigor is the product, not an afterthought.

// 01 - PROBLEM

Most RAG demos stop at "embed some docs, retrieve top-k, stuff the prompt." eve-sage takes the opposite bet: build agentic retrieval with an honest eval harness, on Vercel's brand-new eve framework, using eve's own documentation as the corpus - so anyone learning eve is using an eve agent to do it.

// 02 - APPROACH

  • Agentic multi-hop retrieval: hard questions get decomposed and re-searched until answerable, not one query and hope.
  • Inline citations on every answer, grounded in the eve / AI SDK / Workflow SDK docs.
  • A finite, high-quality corpus makes eval numbers honest and reproducible.

// 03 - ARCHITECTURE

▸ agentic retrieval · CI-gated evals

  1. Question

    The user question

  2. Decompose

    Break hard questions into sub-queries

  3. Retrieve

    Hybrid retrieval over the docs corpus

  4. Rerank

    Rerank retrieved passages by relevance

  5. Answer

    Answer grounded in retrieved sources

  6. Cite

    Inline citations on every claim

  7. Eval

    Eval harness in CI; regressions fail the build

Self-demonstrating corpus
The agent's knowledge base is the documentation of its own stack - self-referential by design, and it keeps the project measurable instead of vibes-based.
Evals in CI, not as an afterthought
Retrieval quality regressions fail the build. The eval harness is the differentiator versus weekend chatbots.

// 04 - PRODUCTION-GRADE

  • Eval harness wired into CI
  • Hybrid retrieval with reranking
  • Tracing for every retrieval hop
  • Built on durable execution with HITL approvals available from the framework

// 05 - ARTIFACTS