AI Assistant — Provider Configuration

The AI assistant answers questions, writes SQL, and suggests plots. It supports several provider backends — from cloud APIs to fully local in-browser inference.

Open Settings (⚙ in the top toolbar) to configure.

Provider overview

Provider	Where it runs	API key required	Works on GitHub Pages
Anthropic (Claude)	Anthropic cloud	Yes	Yes (direct API)
OpenAI	OpenAI cloud	Yes	Yes (direct API)
Google Gemini	Google cloud	Yes	Yes (direct API)
Local OpenAI-compatible	Your machine	Optional	Dev-mode only ‡
In-browser (Qwen2.5-0.5B)	Your browser	No	Yes

‡ Local providers require the Vite dev server for the CORS proxy. See Local provider on GitHub Pages below.

Cloud providers (Anthropic, OpenAI, Google)

Open Settings → AI provider and select the provider.
Paste your API key. The key is stored only in your browser's localStorage.
Click Test to verify the key works.

For Anthropic, the best model for JFR analysis is claude-sonnet-4-6 (the app's default).

Local OpenAI-compatible provider

Connects to any server that exposes an OpenAI-compatible /v1/chat/completions endpoint: llama.cpp llama-server, Ollama, vLLM, LM Studio, or any OpenAI-compatible gateway.

Settings

Field	Description
Base URL	Root URL of the API server — do not include `/v1`.
API Key	Optional. Leave blank for unauthenticated local servers.
Advanced Model	Model name used for chat (e.g. `qwen3:9b`, `gpt-4.1`).

llama.cpp / llama-server (default)

Base URL: http://localhost:8080
API Key:  (leave blank)
Model:    qwen3:1.7b  or  qwen3:9b

Ollama

Base URL: http://localhost:11434
API Key:  (leave blank)
Model:    qwen3:9b  or  llama3.2:3b

Pull the model first: ollama pull qwen3:9b

SAP AI Core / Hyperspace ("Local Hai Proxy") {#sap-hai-proxy}

The SAP AI Core proxy exposes an OpenAI-compatible endpoint. When running locally on port 6655:

Base URL: http://localhost:6655/openai
API Key:  <your-proxy-api-key>
Model:    gpt-4.1   (or gpt-4.1-mini, gpt-5, gpt-5-mini)

Important: The Base URL must end at /openai, not /openai/v1. The client appends /v1/chat/completions automatically.

Available models via the SAP AI proxy:

Model ID	Notes
`gpt-4.1`	Best for complex JFR analysis
`gpt-4.1-mini`	Faster, good for simple queries
`gpt-5`	Latest, highest quality
`gpt-5-mini`	Fast and capable

Quick-start with the SAP proxy in dev mode:

# Set env var so the Vite proxy routes /local-ai-proxy → http://localhost:6655/openai
LOCAL_AI_BASE_URL=http://localhost:6655/openai npm run dev

Then set Base URL to http://localhost:6655/openai and your API key in Settings.

Local provider on GitHub Pages

Summary: The CORS proxy that makes local AI work is part of the Vite dev server. It is not available on the GitHub Pages deployment.

If you want to use a local AI provider with the live web app at https://parttimenerd.github.io/jfr-query/, you have two options:

Option A — Run jfr-query locally (recommended)

# Clone and start the dev server with your local AI proxy URL
git clone https://github.com/parttimenerd/jfr-query
cd jfr-query/core/frontend
LOCAL_AI_BASE_URL=http://localhost:6655/openai npm run dev
# Open http://localhost:3000

This is also the best option for analysing real recordings — you get full JFR file access without uploading data to a third party.

Option B — Use a CORS-enabled endpoint

Some AI gateways and self-hosted servers support CORS. If your local server sends Access-Control-Allow-Origin: * (or the specific GitHub Pages origin), you can enter its URL directly in Settings and the browser will call it cross-origin.

To check: curl -I -X OPTIONS http://localhost:YOUR_PORT/v1/chat/completions If the response includes Access-Control-Allow-Origin, it will work from GitHub Pages.

In-browser model (Qwen2.5-0.5B)

Runs entirely in your browser — no server, no API key, no data leaves your machine. Uses Transformers.js with the onnx-community/Qwen2.5-0.5B-Instruct model (~483 MB, downloaded once and cached).

Enable it in the chat panel by clicking the 🧠 browser routing button, or set AI provider → In-browser model in Settings.

The first message triggers the model download. A progress bar appears in the chat header. After the first load the model is cached in browser storage and loads instantly.

Note: The 0.5B parameter model is suitable for simple questions and SQL suggestions. For complex multi-step analysis, use a cloud or local provider.

Environment variables (dev / self-hosted)

These variables are read at build time by Vite and override the Settings defaults:

Variable	Effect
`ANTHROPIC_API_KEY`	Pre-fill Anthropic key
`OPENAI_API_KEY`	Pre-fill OpenAI key
`GEMINI_API_KEY`	Pre-fill Google key
`LOCAL_AI_BASE_URL`	Enables the CORS proxy for a local AI server (dev only)
`ANTHROPIC_BASE_URL`	Override Anthropic base URL (for custom proxies)

Set them in a .env.local file at core/frontend/.env.local or export before npm run dev.