Elody — Anja Gutierrez

What it is

Elody is the research assistant I actually use. She indexes a markdown knowledge graph, every repository I work in, and my Chrome bookmarks, then exposes the whole thing to Claude Code as 20 MCP tools — search, context loading, code search with project-aware routing, gates and milestones, session history, health checks.

She also runs as a standalone terminal app with an Ink TUI and animated sprites, tracks git commits across every repo, and handles two-way SMS with auto-reply through the same knowledge pipeline. An autonomous /loop engine generates plans and trips a circuit breaker when it stops making progress.

The retrieval problem

Pure semantic search is great at “what did I decide about caching?” and useless at “find the node called rate-limiting” — embeddings smear exact identifiers. Pure keyword search is the reverse.

So retrieval is hybrid, and — this is the part that matters in practice — it has to keep working when the embedding provider is unreachable, out of quota, or the machine is offline. An assistant that returns nothing on a flaky network is an assistant you stop trusting.

The code

Pulled straight from the production codebase — unedited except for trimming.

src/retrieval/search.js javascript

Hybrid retrieval with a real fallback path. The first branch is the load-bearing one: no embeddings available means degrade to keyword search, not fail. When both are available, semantic hits get double weight, keyword scores are normalized into a 0–1 range with score / (score + 5) so an unbounded term count can't dominate a bounded cosine similarity, and the two merge into one ranking.

export function hybridSearch(queryEmbedding, nodeEmbeddings, query, graph, topK = 3) {
  // If no embeddings available, fall back to keyword only
  if (!nodeEmbeddings || nodeEmbeddings.size === 0) {
    return keywordSearch(query, graph, topK);
  }

  const semanticResults = semanticSearch(queryEmbedding, nodeEmbeddings, topK * 2);
  const keywordResults = keywordSearch(query, graph, topK * 2);

  // Merge with weighted scores
  const merged = new Map();

  for (const r of semanticResults) {
    merged.set(r.nodeId, (merged.get(r.nodeId) || 0) + r.score * 2);
  }

  for (const r of keywordResults) {
    // Normalize keyword scores to 0-1 range
    const normalizedScore = r.score / (r.score + 5);
    merged.set(r.nodeId, (merged.get(r.nodeId) || 0) + normalizedScore);
  }

  const results = [...merged.entries()]
    .map(([nodeId, score]) => ({ nodeId, score }))
    .sort((a, b) => b.score - a.score)
    .slice(0, topK);

  return results;
}

Screens

From the running application.

Elody knowledge agent terminal UI — The Ink TUI — search results, session history, and sprite animation.

Want to see more of this one?

The repository is private, but I'm glad to walk through the codebase live — architecture, tradeoffs, the parts that went wrong first.

Email me ← Back to all work