Julian Wiley

Choosing the Right Starter Project in a Local-First Stack

April 11, 2026· 1 min readAgentic Assistants

A practical guide to selecting the right Agentic Assistants starter based on latency, memory, retrieval quality, and operational complexity.

Agentic AssistantsRAGStartersLocal AIArchitecture

Starters Are Product Decisions

examples/starters/README.md lists several starter projects, but the real value is not the list itself. The value is matching each starter to a production pain point.

I use this mapping:

  • local-rag-studio when speed-to-first-answer matters
  • multilevel-storage-kb when retrieval cost and memory tiers matter
  • hybrid-cache-rag-assistant when latency is the hard requirement
  • rag-eval-playground when teams need measurable quality loops
  • repo-intel-hub when source-of-truth is a codebase, not documents

What Changed Recently

The starter set expanded and was formalized through:

  • dedicated README.md per starter
  • config.yaml for each starter
  • an index file in examples/starters/starter_manifest.yaml

This turned "example apps" into reusable starting points with consistent layout.

A Selection Framework That Works

Before I pick a starter, I force these three questions:

  1. Is the bottleneck latency, quality, or governance?
  2. Is the primary corpus documents, code, or mixed assets?
  3. Do I need day-one evaluation metrics or can they come later?

Those answers usually select one starter immediately.

Why This Beats Building From Scratch

Starting from zero encourages accidental architecture.

Starting from a known starter gives:

  • known config shape
  • known runtime flow
  • known extension points

It also improves handoff. Teammates can compare two starter configs much faster than they can reverse engineer one-off glue code.

Practical Takeaway

Treat starter selection like infrastructure selection. Picking the wrong starter is often more expensive than writing custom code later.

If you are unsure, start with local-rag-studio, then migrate to hybrid-cache-rag-assistant or multilevel-storage-kb once you can measure latency and recall tradeoffs.

Related Posts