LLM Configuration

Configure LLM

Documentation index · English · 中文

Article summaries, full-text translation and the daily digest all call an OpenAI-compatible chat-completions endpoint. CFRSS does not ship a model — you bring your own provider, and the key stays yours.

If you skip this page entirely, CFRSS still works: you just don't get summaries, translation or the digest.

---

1. Add a provider

Settings → AI providers → Add, then fill in four fields:

Field	What to put
Name	Anything you'll recognise, e.g. DeepSeek
Base URL	The API root, including /v1 where the provider uses it
Model	The exact model id
API key	The provider's key

CFRSS calls {Base URL}/chat/completions with Authorization: Bearer {API key}. That is the only thing you need to get right.

Base URLs that work

Provider	Base URL	Example model
OpenAI	https://api.openai.com/v1	gpt-4o-mini
DeepSeek	https://api.deepseek.com/v1	deepseek-chat
Moonshot / Kimi	https://api.moonshot.cn/v1	moonshot-v1-8k
Zhipu GLM	https://open.bigmodel.cn/api/paas/v4	glm-4-flash
SiliconFlow	https://api.siliconflow.cn/v1	Qwen/Qwen2.5-7B-Instruct
OpenRouter	https://openrouter.ai/api/v1	openai/gpt-4o-mini

These are examples, not endorsements — check your provider's docs for the current base URL and model names. Anything that speaks the OpenAI chat-completions API works.

⚠️ The endpoint has to be reachable from the public internet. The request is made from your Cloudflare Worker, not from your laptop, so a model running on http://localhost:11434 (Ollama, LM Studio, …) will not be reachable unless you expose it through a tunnel or a public hostname.

2. Assign it

Still in Settings → AI providers, choose which provider handles each function:

Function	Used by
Summaries	The ✦ button in the article toolbar
Translation	The 🌐 button, and the daily digest

You can point both at the same provider, or use a cheap fast model for summaries and a stronger one for translation. Up to 10 providers can be stored.

3. Test it

Each provider row has a Test button. It sends a small request and reports back whether the endpoint, the key and the model name are all correct. Use it before you go looking for the problem somewhere else.

---

What gets sent

When you press Summarize, CFRSS sends the full article text (HTML stripped) to your provider, along with an instruction to produce a short structured summary. Translation sends the article's HTML with an instruction to translate the text while preserving every tag.

Requests carry enable_thinking: false so reasoning models don't spend tokens before answering, and follow your interface language — a Chinese UI asks for a Chinese summary.

Nothing is sent anywhere until you press a button. There is no background summarisation.

Caching

Result	Cached where	Invalidated when
Summary	Per article	Never — recomputed only if you clear it
Translation	Per article and target language	Never
Daily digest	Per calendar day	Regenerated on request

The cache is what makes the second visit instant, and it is also why translating the same article into a second language costs another call.

Translation languages

Ten targets: 中文 · English · 日本語 · 한국어 · Français · Deutsch · Español · Русский · Português · Italiano.

Three display modes, cycled from the toolbar:

• Dual-line (default) — each paragraph followed by its translation
• Side by side — two columns
• Translation only

Your target language and display mode are remembered per device.

Where the key lives

The API key is encrypted with ENCRYPTION_KEY before it is written to the database, and the API never returns it in clear text — the settings page only ever shows a masked value.

That means: if you change ENCRYPTION_KEY, every stored key becomes unreadable and has to be entered again.

Troubleshooting

Symptom	Likely cause
Test fails immediately with a 4xx	Base URL or model name is wrong. Check whether your provider needs /v1.
Test fails with 401 / 403	Wrong API key, or the key has no access to that model.
Test hangs, then times out	The endpoint isn't reachable from Cloudflare's network.
Summaries work, translation doesn't	The two functions are assigned to different providers — check both.
"Maximum of 10 LLM configurations allowed"	Delete an unused provider first.