Selfhosted

@xyro Ah, I see! I’m not using Ollama at the moment — my setup is based on GPT4All with a locally hosted DeepSeek model, which handles the semantic parsing directly.

As mentioned earlier, the pipeline doesn’t just diff pages — it detects new document URLs from the source feed (via selectors), downloads them, and generates structured summaries. Here's a snippet from the YAML config to illustrate how that works:

(extract:
  events:
    selector: "results[*]"
    fields:
      url: pdf_url
      title: title
      order_number: executive_order_number

download:
  extensions: [".pdf"]

gpt:
  prompt: |
    Analyze this Executive Order document:
    - Purpose: 1–2 sentences
    - Key provisions: 3–5 bullet points
    - Agencies involved: list
    - Revokes/amends: if any
    - Policy impact: neutral analysis
)

To keep things efficient, I also support regex-based extraction before passing content to the LLM. That way, I can isolate relevant blocks (e.g. addresses, client names, conclusions) and reduce the noise in the prompt. Example from another config:

processing:
  extract_regex:
    - "object of cultural heritage"
    - "address[:\\s]\\s*(.{10,100}?)(?=\\n|$)"
    - "project(?:s)?"
    - "circumstances"
    - "client\\s*:?\\s*(.{10,100}?)(?=\\n|$)"
    - "(?:conclusions?)\\s*(.{50,300}?)(?=\\n|$)"

Let me know if you're experimenting with similar flows — I’d be happy to share templates or compare how DeepSeek performs on your sources!