Nexus Analyst Demo — Development skill for Claude Code

Nexus Analyst Demo

A side-by-side analytics agent comparison.

Two agents answer the same question over the same BigQuery public datasets. The only thing that differs is **how they retrieve context** before writing SQL.

	agent-nexus/	agent-classic/
Retrieval	One tool: nexus_query over a Nexus context	Three tools: search_schema, search_docs, search_notebooks over a Pinecone index
Index management	None — Nexus handles curation, build, evals	Manual: chunker, embedder, indexer, eval loop
Source files	Uploaded to a Nexus context	Chunked + indexed into Pinecone (see nexus-analyst-demo-ingest)
Tool count	4	6

Same agent loop, same BigQuery executor, same chart generator. The retrieval substrate is the only deliberate difference.

Layout

agent-nexus/                  hero agent — Nexus context, one retrieval tool
  agent/
    connectors/
      nexus.py                Nexus REST client
      bq.py                   BigQuery exec with safety rails
      llm.py                  Vertex AI Anthropic client
    tools.py                  4 tools (1 retrieval + 3 utility)
    agent.py                  run loop + system prompt
    server.py                 FastAPI SSE endpoint
  Dockerfile, pyproject.toml

agent-classic/                baseline agent — Pinecone, three search tools
  agent/
    connectors/
      pinecone.py             Pinecone search wrapper (semantic + rerank)
      bq.py                   (same as agent-nexus)
      llm.py                  (same as agent-nexus)
    tools.py                  6 tools (3 retrieval + 3 utility)
    agent.py                  run loop + system prompt (different retrieval workflow)
    server.py                 FastAPI SSE endpoint (same shape)
  Dockerfile, pyproject.toml

infra/
  deploy_nexus.sh             Cloud Run deploy for agent-nexus
  deploy_classic.sh           Cloud Run deploy for agent-classic

What's deliberately *not* in this repo

• **So