Research teams spend a disproportionate amount of time reading before they start thinking. A team tasked with understanding the competitive landscape in a new market might read 50 reports, 200 papers, and hundreds of news items before synthesizing a view. Most of that reading is triage — figuring out what is relevant and what is not.
Crucible handles the triage. The analyst handles the judgment.
The Pattern
The most common research synthesis pattern using Crucible is a two-pass pipeline.
In the first pass, each document is sent individually with a prompt designed to extract the specific information relevant to the research question. The output is a structured summary per document — key claims, supporting evidence, and a relevance score.
In the second pass, the structured summaries are batched into a single request asking the model to synthesize across them. At 128K token context, you can fit summaries of 30-50 documents in a single synthesis request.
What Makes It Work
The quality of the synthesis depends heavily on how the first pass is structured. Vague extraction prompts produce vague summaries, which produce vague syntheses. The more precisely you define what you are looking for in the first pass, the more useful the synthesis becomes.
Specifying an output format for the first pass — a JSON object with defined fields rather than free text — makes the second pass significantly more reliable.
Research teams spend a disproportionate amount of time reading before they start thinking. A team tasked with understanding the competitive landscape in a new market might read 50 reports, 200 papers, and hundreds of news items before synthesizing a view. Most of that reading is triage — figuring out what is relevant and what is not.
Crucible handles the triage. The analyst handles the judgment.
The Pattern
The most common research synthesis pattern using Crucible is a two-pass pipeline.
In the first pass, each document is sent individually with a prompt designed to extract the specific information relevant to the research question. The output is a structured summary per document — key claims, supporting evidence, and a relevance score.
In the second pass, the structured summaries are batched into a single request asking the model to synthesize across them. At 128K token context, you can fit summaries of 30-50 documents in a single synthesis request.
What Makes It Work
The quality of the synthesis depends heavily on how the first pass is structured. Vague extraction prompts produce vague summaries, which produce vague syntheses. The more precisely you define what you are looking for in the first pass, the more useful the synthesis becomes.
Specifying an output format for the first pass — a JSON object with defined fields rather than free text — makes the second pass significantly more reliable.
Research teams spend a disproportionate amount of time reading before they start thinking. A team tasked with understanding the competitive landscape in a new market might read 50 reports, 200 papers, and hundreds of news items before synthesizing a view. Most of that reading is triage — figuring out what is relevant and what is not.
Crucible handles the triage. The analyst handles the judgment.
The Pattern
The most common research synthesis pattern using Crucible is a two-pass pipeline.
In the first pass, each document is sent individually with a prompt designed to extract the specific information relevant to the research question. The output is a structured summary per document — key claims, supporting evidence, and a relevance score.
In the second pass, the structured summaries are batched into a single request asking the model to synthesize across them. At 128K token context, you can fit summaries of 30-50 documents in a single synthesis request.
What Makes It Work
The quality of the synthesis depends heavily on how the first pass is structured. Vague extraction prompts produce vague summaries, which produce vague syntheses. The more precisely you define what you are looking for in the first pass, the more useful the synthesis becomes.
Specifying an output format for the first pass — a JSON object with defined fields rather than free text — makes the second pass significantly more reliable.