While the UI presents "Search" and "Question" as distinct modes, the underlying operational difference is one of **query transformation complexity and result synthesis.** This isn't merely a UI toggle; it's a shift in how Elicit's language model interprets your input and structures its output, with significant implications for latency and the cognitive load required to parse the results.
A **Search** is essentially a sophisticated keyword-based retrieval operation. You provide terms or a phrase, and Elicit performs a semantic search across its corpus (primarily the Semantic Scholar database) to return a list of relevant papers. The LLM's role here is largely to re-rank and summarize the *abstracts* of the top matches. The output is a list.
```json
// Conceptual "Search" flow:
User Input: "BERT transformer attention mechanisms" -> LLM: [Generate search query embeddings] -> Vector DB Retrieval -> LLM: [Summarize each abstract] -> Return List of Paper Summaries.
```
A **Question**, however, initiates a multi-stage reasoning pipeline. You ask a direct, answerable question. Elicit must then:
1. Decompose the question into potential search queries.
2. Retrieve relevant papers (similar to Search).
3. **Extract specific claims, data, or conclusions from the *full text* of those papers** that pertain directly to the question.
4. Synthesize these extractions into a coherent, evidence-backed answer.
```json
// Conceptual "Question" flow:
User Input: "Does BERT outperform GPT-3 for sentiment analysis?" -> LLM: [Decompose to sub-queries] -> Vector DB Retrieval -> LLM: [Extract relevant sentences from full texts] -> LLM: [Synthesize extractions, note conflicts] -> Return Answer with Citations.
```
The performance and latency profile differs markedly:
* **Search Latency:** Governed by retrieval + abstract summarization. Lower computational cost, faster response.
* **Question Latency:** Adds the costly steps of full-text extraction and synthesis. Higher computational cost, significantly slower response. The quality is highly dependent on the specificity of the question and the availability of directly relevant text in the retrieved papers.
Therefore, the pragmatic distinction for a user focused on efficiency:
* Use **Search** when you are conducting exploratory research, need a broad overview of a field, or want to quickly gather a set of potentially relevant papers for later manual review. It's a high-recall, lower-precision tool.
* Use **Question** when you have a specific, fact-based hypothesis and require a synthesized answer drawn from the literature. This is a high-precision tool, but you pay for it in processing time and must carefully vet the provided citations, as the extraction step can sometimes pull sentences out of context.