Architecture

How a research campaign runs, and what leaves your machine by which door, with the file of the code behind each line.

How a research campaign runs

manifest.json
     │  campagne.py
     ▼
SearXNG  ── queries
     ▼
light HTTP fetch  ──►  real browser (only if the page does not suffice)
     ▼
collecte.jsonl  +  operations.jsonl
     ▼
OpenCode  ── ranking, then the report
     ▼
rapport_<timestamp>.md

The manifest is a JSON file of targets: a query for SearXNG, a fixed url, a produit (find the one page that really is the product page), or a table_reference (a list of trusted domains for a subject). Only the campaign identifier and the targets are required.
Discovery. A query target goes to the SearXNG instance you configured, which in turn queries other search engines that Dinoer does not control.
Collection. Each page is fetched with a plain HTTP request, after its robots.txt has been read. Only a page whose text cannot be read that way goes to a real browser, one Playwright process per page. A target that fails does not stop the others.
The corpus. One collecte.jsonl per campaign, one line per page retained, and operations.jsonl, which records every attempt, failed or not.
The report. The pages are reordered (by time window and by local embeddings, when the manifest asks for it), cut to a size budget, and handed to OpenCode, which writes the body of rapport_<timestamp>.md. Dinoer appends the list of sources itself; the model never writes it.

shot.py and rpa.py work differently: each call is its own process and ends as soon as it has answered. Why →

What leaves your machine, and by which door

What	Where it goes	Who decides	In the code
Discovery queries	the SearXNG instance you configured, which queries other engines in turn	you, with `DINOER_SEARXNG_URL`	`lib/searxng.py`
Page requests	each site directly, its `robots.txt` read first	the targets in your manifest	`lib/fetch_leger.py`, `rpa.py`
Collected text	the OpenCode model you name, for the report, for targeted extraction, and to choose a product page (the first 1,500 characters of each candidate)	you, with `DINOER_OPENCODE_MODEL`	`lib/modeles.py`, `lib/synthese.py`, `lib/extraction.py`, `lib/selection_candidats.py`
A subject, for a reference table	the same OpenCode model, the subject only, when no table exists for it yet	the `table_reference` target	`lib/tables_reference.py`
Embeddings	your local Ollama, `http://localhost:11434`: the text stays on the machine unless you point `DINOER_OLLAMA_URL` elsewhere	you	`lib/vector.py`
A notification	a ntfy server, `https://ntfy.sh` unless you set another: the campaign identifier and the number of sources. Only when a ntfy `topic` is set	you, with `ntfy_topic`, `DINOER_NTFY_TOPIC`, `DINOER_NTFY_URL`	`campagne.py`, `lib/ntfy.py`
Your credentials	nowhere: resolved inside the Playwright process, from the encrypted directory	—	`lib/repertoire_chiffre.py`

Source: the code of version 1.0.1, read on 27 September 2026. The models listed under opencode/ are hosted by their provider (opencode.ai/zen, according to opencode models --verbose, OpenCode 1.18.32). What the model may do with the web is on the trust page.

Architecture

How a research campaign runs

What leaves your machine, and by which door

In detail