CHLLM

How it works

How CHLLM works, the levels, the modes — and how to get the most out of your researches.

What CHLLM does

CHLLM doesn't answer from memory. For every question, it runs an investigation: it searches the web, opens the documents, reads them, cross-checks what they say, and writes a report where every claim points to its source. Every sentence can be verified in one click.

That's the difference with a conversational assistant: an answer takes time — from twenty minutes to an hour depending on the level — but it is documented. You can pass it on, challenge it, cite it. A link that could not be verified is flagged as such.

How a research works

A research mobilises a team of models, each at its post. A supervisor analyses your question and breaks it into research angles. Researchers query the web and read the sources, one by one. The supervisor re-reads their notes, cross-checks the information and makes sure nothing is missing from the answer — as long as a point remains open, it launches another search. A writer then drafts the report.

“Help me frame it” — asking the right question

The quality of a research is decided before it starts: a badly framed question yields a report that misses the point. So CHLLM offers to frame first — it's the default, and it's free.

You describe your topic freely. The Assistant does its homework — it consults real sources during the framing — then asks you precise questions, one at a time, with clickable answers. It then writes the consolidated question into your field, where you can edit everything, and sets up the research, justifying each choice: the source perimeter, the level, the cost.

Nothing is charged during framing, and the research only starts when you launch it. “Direct research” remains available to start immediately with your question as it is.

Choosing the level (Éco, Standard, Profond)

All three levels follow exactly the same protocol. What the level chooses is the intelligence at each post: the more powerful the models, the finer the analysis — the question better decomposed, the sources better understood, the contradictions better caught.

Éco
When the difficulty is gathering, not judging: regulation, procedure, state of play, figures. Solid models at every post, with the writing handed to one of the most powerful. Complete, sourced, well written.
Standard
The supervisor and the researchers step up in power: more discernment in decomposing the question, weak signals, disagreements between sources. Whenever the answer calls for interpretation.
Profond
The most powerful models everywhere, and a wider exploration: more angles in parallel, more searches and reads per angle. For multi-part briefs and topics where sources contradict each other.

Modes

A mode defines the source scope and the behavior of a research. Pick it just before launching.

Five default modes are maintained for everyone:

Général
The open web. No domain restriction.
Droit français
Légifrance, Cour de cassation, Conseil d'État, Juricaf, EUR-Lex. Every citation comes from official French and European legal sources.
Académique francophone
Cairn, HAL, Persée, theses.fr, OpenEdition, Érudit. Peer-reviewed articles and preprints from the French-speaking academic world.
Scientifique international
PubMed, arXiv, bioRxiv, medRxiv, Nature, ScienceDirect, PLOS, Springer, Wiley, DOI. Queries are issued in English to reach the international literature; the report is written in the language you choose.
Médical & santé
HAS, WHO, EMA, ANSM, Vidal, Cochrane, PubMed, CDC. Official guidance and meta-analyses.

The promise: in any mode other than Général, every citation in the report comes from the announced perimeter. If an exception slips through (the model can occasionally fabricate a URL or cite a page mentioned inside a perimeter source), it's marked with a ⚠️ and surfaced at the bottom of the report.

You can also duplicate a mode to adapt it, or hide it.

Official databases queried at the source

Beyond web search, the researchers have dedicated access to some fifty official databases and registers. Whenever the topic calls for it, the text or figure is taken directly from its producer, cited with its official URL, then verified after writing.

The “?” bubble next to the mode, in the composer, sums up what the chosen mode favours.

Your custom modes

Go to Settings → Research modes. You can create a mode from scratch (“New mode”), or click “Duplicate & customize” on a default mode to start from a base. Your modes belong only to you.

Each field in the editor has a small ? bubble: hover it (or tap it on mobile) for a quick reminder. In detail:

Label
The name shown in the search picker.
Search API
Keep tavily in the vast majority of cases: it's the only one that lets you filter by domain (the two fields below). The openai / anthropic options use the model's native web search (no domain filtering), and none disables web search.
Display order
The mode's position in the picker — the smaller the number, the higher it appears.
Enabled
Uncheck to remove the mode from the picker without deleting it (handy for a draft or a mode set aside).
Clarification
Allow or not scoping questions for this mode.
Allowed domains
Limit the research to these sites only (one per line, e.g. legifrance.gouv.fr). Leave empty to allow the whole web. Both domain fields only work with the tavily engine: they're automatically greyed out if you pick another engine.
Excluded domains
Sites to keep out of the results (one per line).
Source instructions
Free-form instructions added to the research, e.g. “prefer primary sources and case law, ignore blogs”.

Tip: create one mode per type of work (one for labor law, one for market watch…), then you just pick it.

Framing your request well

CHLLM follows your instructions in plain language: write them straight into your question. Some requests are honored reliably, others stay indicative. Better to know upfront.

What the engine honors

Report shape
“as a comparison table”, “as a bulleted list”, “a single synthesis section”, “comparison between A and B”, “by sections: X, Y, Z”.
Source type
“official / primary sources”, “academic / peer-reviewed sources”, “avoid blogs and SEO sites”. To restrict to specific sites, use a mode.
Source language
“French sources only”, “English sources”.
Angle, audience, focus
“for a busy executive”, “legal angle”, “risk-focused”, “plain-language”, “operational perspective”.
Report language
via the dedicated selector (quotes stay in their original language).

⚠️ What the engine does NOT guarantee

While it runs

The research runs in the background: you can close the tab, a notification tells you when the report is ready (enable them in Settings). An estimated delivery time is shown while you wait. One research at a time: if one is already running, the home screen offers to reopen it.

An interrupted research — cancelled, or stopped by an incident — can resume where it left off: the work already done is neither lost nor billed twice.

Documentary mission: your documents and the Pack (invite-only Beta)

On invited accounts the composer becomes a four-step mission — Mission, Sources, Deliverables, Validation — freely navigable: Validation always recaps what will be read, produced and charged before launch.

Your documents
At the Sources step: upload files (PDF, Word, PowerPoint, Excel, text, Markdown, CSV) from your device or import them once from Google Drive (you pick the files, nothing else is read). They join your library, reusable across missions. The researchers read them page by page, cite them like any source — the link opens the cited page, visible to you only — and confront them with public sources. A scanned document without text is flagged as unreadable.
Instructions on sources
A free instruction, frozen at launch, passed to the supervisor and the researchers: “confront my documents with the official texts”, “flag contradictions”.
The deliverables Pack
At the Deliverables step: on top of the report, the executive summary, the action plan, the questions for a professional adviser, the sources index (with verification verdicts). Produced after delivery, derived from the verified report and checked against it; each deliverable can be read online and downloaded as PDF, Word or Excel, the ZIP bundles everything.
Confidentiality
Your documents stay tied to your account, are never shared nor reused; a public share page never shows their content. You can delete them at any time: the citation stays in the report, its content is no longer available.

Reading and using a report

The report opens with a table of contents and ends with the numbered list of sources; every claim is cited inline. Links the engine could not verify are flagged. Your reports stay in the History; you can favourite them, share them by public link — with or without your original question — or export them as Markdown and PDF.

Credits

CHLLM runs on credits, with no subscription. The cost of an analysis depends on the level and is shown before every launch. Framing a request is on us five times when you open your account; after that it costs one credit, deducted from the price of the analysis that follows. A failed analysis is not charged. You top up with packs, no subscription.

Installing the app

CHLLM installs like an application, without going through a store.

On a computer (Chrome or Edge — Mac, Windows, Linux)
Open the site, then click the install icon on the right of the address bar (a screen with an arrow). If it doesn't show: menu ⋮ → “Install CHLLM” (depending on the version: “Cast, save and share” → “Install page as app”). The app then opens in its own window, with its icon in the Dock or the taskbar.
On a Mac with Safari
File menu → “Add to Dock”.
On iPhone / iPad
Open the site in Safari, tap Share (the square with an arrow), then “Add to Home Screen” and Add.
On Android
Open the site in Chrome, menu ⋮ → “Install app”, confirm.

Once installed, enable notifications in Settings to be told as soon as a report is ready.

Prepare a request — framing on us Five framing sessions when you open your account — no credit card. You only pay for the analysis, once the scope is approved.
← Back to CHLLM