Quickstart

Summary

Five minutes from a running stack to a seeded cluster and the chat agent. Assumes you've already done the Install step (make up + make migrate).

1. Seed a cluster and sample data

RelyLoop ships a 1,000-document sample products index so you can run a real study without wiring up your own corpus first.

make seed-clusters   # register local-es + local-opensearch
make seed-es         # index samples/products.json into local-es (1,000 docs)

make seed-clusters inserts two rows into the clusters registry — one for the bundled Elasticsearch, one for OpenSearch. make seed-es loads the sample catalog so there's something to search.

2. Open the chat agent

open http://localhost:3000/chat

The conversational agent is the front door. It describes the loop, proposes a search space, and dispatches the same tools the API exposes — start_study, generate_judgments_*, open_proposal. Ask it something like:

Run a study against the products index on local-es and tune relevance for my query set.

3. Watch the loop run

A study spins up an Optuna TPE optimization over the query-time search space. Each trial renders the query templates with a candidate parameter set, runs your query set against the cluster, and scores the results against your judgments with ir_measures. The agent reports progress; the /studies and /studies/[id] pages show trial scatter plots and parameter importance.

4. Review the proposal

When the study finishes, RelyLoop writes a digest (a plain-language summary of what moved the metric and why) and stages a proposal — the winning configuration, ready to open as a Pull Request against your config repo. Review it on /proposals.

What just happened

flowchart LR
    A[Seed cluster<br/>+ sample data] --> B[Chat agent<br/>proposes search space]
    B --> C[Optuna TPE<br/>runs trials]
    C --> D[ir_measures<br/>scores each trial]
    D --> E[Digest + proposal]
    E --> F[Pull Request]

You ran the full loop end-to-end against bundled data. Next, do it with a real example and a real PR in Your First Optimization Loop.

Prefer the guided tutorial?

The first-study tutorial walks the entire path from git clone to "PR opened in GitHub" with screenshots, including a local-LLM (Ollama) variant.