Three Ways to Talk to NataPulse: Chat Modes Explained

Most AI chat products blur an important line: sometimes the answer comes from a model’s general knowledge, sometimes from a retrieval layer, and the interface rarely tells you which. NataPulse’s chat surface, live since late June, takes the opposite position. When you open /chat you choose one of three explicit paths — a direct model conversation, a cited search over NataPulse’s curated intelligence, or a hand-off to the full Deep Research pipeline — and the product never infers the path from the wording of your message. The mode is a setting you select, not a guess the system makes. And, as everywhere on NataPulse, every path produces research evidence, never trade instructions.

Mode one: LLM — a model conversation, honestly labeled

The first path is the plainest: your question goes straight to a language model. No retrieval runs, no documents are fetched, and the answer carries zero citations — by design, and visibly so. The response arrives with an empty citations list and zeroed source counters, so there is no ambiguity about what you are reading: fluent model prose, drawing on the model’s training, with all the strengths and known failure modes that implies.

That honesty is the feature. Plenty of questions do not need curated market evidence — you may want to rephrase a paragraph, sanity-check a definition, or think out loud. What you should not get in those moments is an answer dressed up with borrowed authority. In LLM mode, NataPulse makes no claim that the answer is grounded in its data, because it is not.

Mode two: Search NataPulse — cited answers from curated evidence

The second path flips the contract. In Search NataPulse mode, your question runs against the platform’s curated intelligence: published market events from the ingest pipeline, reports you are authorized to read, and public memory. The answer comes back with numbered [N] citation markers placed directly in the text, and a Sources panel alongside the conversation shows the evidence cards behind each marker. Selecting a marker highlights the corresponding evidence, and each citation stays labeled by type — event, report, memory, or permitted web source — so historical internal knowledge is never confused with fresh external evidence.

The retrieval scope is deliberately conservative. Search mode reads only published events and public, curated memory — never drafts, never another workspace’s data. A cited answer can still be incomplete or wrong, and the documentation says so plainly: the right habit is to check whether the citation supports the exact claim, whether the evidence is current, and whether counterevidence is missing. But the failure modes are inspectable, which is the difference between an answer you can audit and one you have to take on faith.

Mode three: Deep Research — from a question to a dossier

Some questions outgrow a chat reply. The third path is an explicit action that hands your question to NataPulse’s Deep Research pipeline — the multi-agent investigation workflow that plans, gathers, debates, and synthesizes across specialist roles, streaming its progress as it works. Instead of a paragraph, you get a run: a structured dossier organized around the fixed 51-section Master Index, with per-section coverage status, claim-level citations on material numbers, and a confidence score derived from coverage, citation, and freshness rather than asserted by the model.

Crucially, this hand-off never happens silently. Pressing send in a chat thread does not quietly escalate into a research run; Deep Research starts only when you invoke it. The interface can detect when a question looks like it deserves a deeper investigation and offer a Deep Research action with a proposed scope — but offering is as far as it goes. The decision, and the wait, remain yours.

The design principle: no silent substitution

The three modes exist because they are three different epistemic contracts. Uncited model prose, cited curated evidence, and a full multi-agent dossier are not interchangeable goods at different speeds; they make different promises about where the words came from and how much you should lean on them. NataPulse’s position is that the user should always know which contract they are in — and that the product must never trade one for another behind the scenes.

That principle extends to failure. If a path is unavailable — say, the model backend for LLM mode is not configured — the product returns an honest error rather than substituting something else, such as passing off a retrieval digest as a model answer. A degraded answer wearing the wrong label would be worse than no answer, because it corrupts the one thing the mode selector is supposed to guarantee: knowing what you are reading.

None of this makes any single mode “the good one.” LLM mode is the right tool for unencumbered thinking, Search for auditable answers about what NataPulse has actually observed, Deep Research for questions worth a dossier. The point of the design is smaller and firmer: whichever you choose, the label on the answer is true.