Setup & FAQ
For version 1.5.0. This page assumes you are comfortable with servers, ports and API keys — if that is not you, the quick start guide covers the same setup in plain language.
Choosing a provider
Any of these work. The only real decision is whether the review text leaves your computer.
| If you want… | Use |
|---|---|
| Nothing to install, nothing leaves your machine | Built-in — the on-device model in Chrome, on capable hardware |
| Nothing to leave your machine, free | Ollama, or any local server below |
| The simplest local setup | Ollama — one install, one command |
| A GUI for managing local models | LM Studio or Jan |
| Best quality without local hardware | OpenAI, Anthropic or Gemini |
| Fast and free-tier hosted | Groq or Gemini |
Built-in model
The Built-in provider uses the on-device model that ships with Chrome, through its Prompt API. The extension makes no request on this path — the reviews are processed by Chrome on this device — and Chrome downloads the model once (a few GB). A new install starts on it wherever it is available; saved settings are never switched over.
- Chrome 138 or later on Windows 10/11, macOS 13+, Linux or ChromeOS Plus — not on phones
- About 22 GB free on the drive holding the Chrome profile. Below 10 GB Chrome deletes the model again, and it downloads anew on the next use
- A GPU with more than 4 GB of VRAM, or 16 GB of RAM and 4 CPU cores. CPU-only machines work but can take minutes per summary; the request gives up after 5 minutes
- An unmetered connection for the one-time download
Check availability in Settings → Built-in asks Chrome directly. When the
model is supported but not on disk yet, Download now fetches it ahead of
the first analysis; otherwise the first Analyze starts the download and collects
reviews while it runs. chrome://on-device-internals shows the model's state,
download progress and any error from Chrome.
It writes English, Japanese, Spanish, French and German. With another output language chosen, the summary is written in English and the result says so; Auto follows the reviews. Its context window is small and not published, so the prompt is measured against it at run time and trimmed to fit — expect fewer reviews analyzed than with a server-hosted model, and the result header gives the real count.
Local setup
Ollama
Install Ollama, then:
ollama pull llama3.2— or any model you preferollama serve
In settings choose Ollama, press Test connection, and the model field will autocomplete from the models you actually have installed. Running it on another port or machine? Change Server endpoint.
LM Studio, llama.cpp, Jan, vLLM and friends
Start your server with its OpenAI-compatible API enabled, then in settings pick Local server and choose your runtime from the preset list — the address is filled in for you:
| Runtime | Default address |
|---|---|
| LM Studio | http://localhost:1234/v1 |
| llama.cpp server | http://localhost:8080/v1 |
| Jan | http://localhost:1337/v1 |
| vLLM | http://localhost:8000/v1 |
| KoboldCpp | http://localhost:5001/v1 |
| text-generation-webui | http://localhost:5000/v1 |
| LocalAI | http://localhost:8080/v1 |
| Ollama (OpenAI-compatible API) | http://localhost:11434/v1 |
Press Test connection — the model list loads from the server itself.
Hosted API setup
Pick the provider, paste its API key, press Test connection. That check uses the provider's free model-list endpoint, so it validates your key without spending any tokens.
Once saved, a key is encrypted and never shown again — the field displays a ✓ Saved badge. Leave it blank to keep the stored key; use ✕ to remove it.
Settings worth knowing
Review scope
- All — everything collected, in Google's own relevance order
- Recent — sorts the page by newest first, then takes the newest N
- 1 / 3 / 6 / 12 months — sorted newest-first and filtered by date
Analysis depth
How many collected reviews actually reach the model: Quick (~100), Balanced (~250) or Thorough (~500). More is slower, and on a local model needs a bigger context window. The result header always tells you the real number, for example "113 of 1,043 reviews analyzed".
Reviews are not simply truncated — every 1–2★ review is kept, since complaints are the scarcest signal, and the rest are sampled evenly across the whole time range.
Context window (local models)
Set this to match what your model actually supports. If it is too small for the chosen depth, the extension automatically trims the prompt to fit rather than overflowing it, and says so in the service-worker console.
Troubleshooting
"Ollama is not ready" / "Cannot reach Ollama at…"
The provider is checked before any reviews are collected, so this appears within a second and
nothing was scraped. The server is not running, or is on a different address: start it with
ollama serve, or correct Server endpoint in settings, then press
Test connection to confirm.
The same message with a provider name in it — "Gemini is not ready" — means that provider refused the check, usually a wrong or expired API key. Analyze anyway is there for the uncommon setup that fails the check but works regardless: the check calls the provider's model-list endpoint, and a self-hosted server can serve chat without offering one.
"Built-in model is not ready"
Chrome reports that this computer cannot run its on-device model, so nothing was scraped.
Usually the hardware or free disk space falls short of the list above, Chrome is older than
138, or a policy (GenAILocalFoundationalModelSettings) turns the model off.
chrome://on-device-internals says which. Choose another provider in settings.
On a new install where this happens the extension starts on Ollama instead, and its error
says that no provider is set up yet.
If the model is available but the download never starts, open Settings → Built-in and press Download now: Chrome starts the download only from a click, and the Analyze click can expire before the model is requested.
My custom endpoint will not connect
- Include the version path — usually
/v1, not just host and port - If it is not on
localhost, Chrome will ask permission when you save. Accept it - Confirm the server was started with its OpenAI-compatible API enabled
The summary is in the wrong language
Set Output language in the Analysis section. Auto follows the reviews' own language; pick a specific language to pin it.
The time filters seem to be ignored
Review dates are read from the page text, and only English relative dates ("3 months ago") are understood. On a non-English Maps interface most dates cannot be parsed — those reviews are kept rather than silently dropped, and the result screen warns you when it happens.
It is slow
- Lower Analysis depth to Quick
- Reduce Maximum reviews to collect
- Use a smaller local model, or a hosted provider like Groq
- Press ⚡ Summarize Now mid-collection to stop gathering and summarize what is in hand
I want a fresh result, not the cached one
↺ Re-analyze reuses the collected reviews and just re-runs the AI. ⟳ Fresh re-scrapes the page from scratch.
What it does when you are not using it
Nothing. There is no background activity, no scheduled work and no connection kept open. The extension's worker wakes only to make the request you asked for, and the content script on a Maps page does nothing until the popup tells it to. Close the popup mid-analysis and the run is abandoned — the result is discarded, though the page may finish the scroll already in progress. Cancel stops it properly.
Still stuck
Open an issue on GitHub or email minglau0315@gmail.com. Including your provider, model and anything in the service-worker console helps a lot.