Frame a Question a Machine Can Help With

Two research tools publish query advice on their own help pages, and the advice is incompatible. Elicit's page on improving search results lists language model long context as a query to avoid. Perplexity's Search API best practices recommends queries of exactly that shape and calls the alternative vague.

Neither is wrong. They are describing different retrieval systems, and the useful takeaway is that there is no such thing as a well-formed research question in general. There is only a question well formed for the system you are about to hand it to. So decide the shape first.

Elicit prints the good and bad version of the same query

Elicit runs semantic search, described in its help centre as search "based on the meaning of your query to find papers," and it wants the meaning to be present. Two worked pairs from that page:

Good: "What techniques work best to allow a language model to accept longer context?"
Avoid: "language model long context"

Good: "How much does mass deworming improve health?"
Avoid: "deworming effectiveness"

Look at what the good versions carry that the keyword versions do not. A verb with a direction ("improve"), a population or scope ("mass"), and an implicit outcome measure ("health"). The keyword string has a topic. The question has a claim you could be wrong about.

The same page gives three repair moves when results disappoint, and they map to three different diagnoses. Broaden the query "to cast a wider net if the results aren't relevant at all." Narrow it "if you're getting lots of semi-relevant results but no highly-relevant results." And "Clarify the language of your query, if Elicit seems not to understand it." Those are not interchangeable. Broadening a query that already returns semi-relevant noise makes it worse.

Perplexity's search docs recommend the shape Elicit warns against

Now the other system. Perplexity's Search API documentation contrasts two queries directly. Better: artificial intelligence medical diagnosis accuracy radiology. Avoid: AI medical. Elsewhere it offers artificial intelligence machine learning healthcare applications 2024 in place of a bare search for AI, with the principle stated as "Specific queries with context, time frames, and precise terminology yield more relevant and actionable results."

The recommended form is a keyword stack. It is precisely the shape Elicit tells you to replace with a sentence. One caveat on scope, since it matters for whether this applies to you: that guidance covers Perplexity's Search API, which is the surface where the company publishes explicit query-shape rules. It is not necessarily advice for typing into the consumer answer box.

The practical rule that falls out of the disagreement is short. If the system does semantic retrieval over abstracts, give it a sentence with a claim in it. If the system does keyword or hybrid retrieval over the open web, give it the four or five terms that would appear in the document you want. Do not carry one habit into the other tool and blame the tool.

The shape of the question decides which tool can answer it at all

Before phrasing, sort the question into one of three buckets, because each has a different failure mode.

Questions about a body of literature you have not read yet. This is what Elicit's workflow picker is organised around, and the workflows are genuinely different tools rather than presets: Find Papers, Research Report, Systematic Review, Extract Data, Chat with Papers, and Research Agent sessions. The guidance on choosing is that it "depends on what stage your research is in, how granularly you want to control and interact with the results, and what output you're looking for Elicit to provide." A question you cannot yet name the outcome measure for belongs in Find Papers, not in a systematic review.

Questions with a quantitative filter baked in. Consensus searches, in its documentation's phrasing, "over 200 million peer-reviewed academic research papers," and its search tool takes parameters rather than expecting you to say everything in prose: study_types, year_min and year_max, a human filter, and medical_mode for clinical questions. If your question contains the words "in humans" or "randomised trials only," those are parameters, not adjectives.

Questions about documents you already have. NotebookLM, which Google's help pages now label Gemini Notebook while still hosting them under the notebooklm path, is explicit about its boundary: it "answers questions based on the information provided in your uploaded sources. If the answer isn't in the source material, it won't provide a response." That is a feature for close reading and a wall for discovery. Asking it what the literature says is asking it what your folder says.

Constraints belong in the query, not in the follow-up

The most common wasted hour in AI-assisted research is running a broad search, reading twenty summaries, and only then remembering the constraint that would have excluded eighteen of them.

Write the constraints down before the first search. Elicit exposes filters for publication year, study type, keywords, and a "Has PDF" filter for papers with an available full text, and it recommends keyword filters specifically for the case where "Elicit seems to be ignoring part of your query." That last line is worth internalising, because it names the failure honestly: semantic search can silently drop a clause it does not weigh heavily, and a keyword filter is how you force it back in.

Volume constraints matter too if the answer lives in your own documents. NotebookLM's published limits are "100 notebooks, with up to 50 sources each and 500,000 words each," and "The current limit is 500,000 words per source or up to 200MB for local uploads." Fifty sources is a decision about which fifty, made in advance.

On the search side, Perplexity's API lets you "Pass up to five related queries in one request to explore different angles of a topic. Each query is processed independently." The billing note attached is a small, useful detail: a multi-query request "is billed as one Search API request, while rate limiting counts each query in the array." Five angles on one question, priced as one.

Give the tool permission to come back empty

A question phrased so that only one answer fits will get that answer. Perplexity's prompt guide for grounded agents makes the countermeasure an explicit instruction: "If searches do not return relevant results after trying alternative phrasings, say so explicitly rather than providing speculative information." It pairs that with a rule for near misses: "If you find related but non-matching results (for example, a different year, a parent company, or a subsidiary), state the mismatch explicitly before answering."

Both of those are things you can write into your own question, in any tool, whether or not it has an instructions field. Ask what the strongest evidence against the claim is, in the same session, before you have read the evidence for it.

The guide adds one constraint on how much you write: instructions are "re-read on every turn of the agent loop, so bloat compounds across tool calls." A page of caveats is not free.

There is one job none of these tools will do for you, and Elicit is the only vendor here that says so in plain words on its own limitations page: "Elicit does not yet know how to evaluate whether one paper is more trustworthy than another," and it "summarizes the findings of a bad study just like it summarizes the findings of a good study." The quality bar has to be in your question, as a study-type filter or a stated exclusion, because the ranking will not supply one.