Quartet Auto Search and Concept Search Best Practices

General AI Searching Disclosure Optimization

Well-crafted disclosures significantly improve Quartet Auto Search and Concept Search (Quartet AI searching) performance. The system relies on descriptive, context-rich input to identify relevant technical material, making the quality and clarity of the text a critical factor in retrieval accuracy.

Use Descriptive Text Input

Quartet AI performs best when the input reads like a clear, descriptive technical explanation rather than a minimal query. Instead of thinking in terms of keywords, it is helpful to approach the input as a short disclosure describing what needs to be found. The more detail included, the easier it is for the system to understand both the invention itself and the surrounding technical space.

• Enter text that directly describes the invention, system, or concept of interest.
• Favor completeness over brevity—more descriptive input generally leads to stronger results.
• Include technical detail that would typically appear in a patent specification.

Add Context to Guide the Search

A strong starting point is simply writing out the concept in plain technical language. Expanding on that description tends to improve results. For example, restating the same idea in different ways can help capture how similar concepts might be described across various documents. Defining important terms also helps anchor the description in the correct technological domain, especially when dealing with specialized or ambiguous language.

Generative AI can help improve Quartet AI inputs by rewriting a disclosure in multiple ways, introducing varied phrasing and terminology that increases the chances of matching across different documents.

However, care should be taken when using public AI tools. Many process data externally, which can create confidentiality risks for sensitive or unpublished material. Always follow data handling policies and use approved, secure systems where possible.

• Restating features in different terms helps capture semantic variation.
• Improves recall across documents that use different terminology.
• Clarify specialized terminology or ambiguous language.
• Explain in more detail how components operate or interact, or what the invention may be used for, include practical use cases or operational scenarios.
• Generative AI may be used to introduce varied phrasing and terminology, which increases the likelihood of matching across different documents.
• Public AI tools should be used with caution, many process data externally; this can create confidentiality risks for sensitive or unpublished material.
• Always follow data handling policies and use approved, secure systems when working with proprietary information.

Include Relevant Background

Context plays a critical role throughout this process. Without enough of it, results may drift into adjacent or unrelated fields. Providing background information helps ensure the system is looking in the correct area, but it is important to keep that background tightly focused. Including unrelated features or describing other inventions that are not central to the core idea can dilute the search and reduce overall relevance.

• Include enough context to define the technical field and problem space.
• Provide supporting detail that clarifies how the invention fits within existing technology.

However, avoid introducing unrelated concepts:

• Do not include descriptions of inventions or features that are not central to the inventive idea.
• Irrelevant background can dilute feature matching and negatively impact scoring.

Emphasize Clarity and Relevance

In general, clearer and more descriptive input leads to better outcomes. A well-written disclosure that defines the technology, explains how it works, and situates it within the right context allows Quartet AI to more accurately identify relevant documents and prioritize the strongest matches.

• Keep descriptions aligned with the core inventive concept.
• Use clear, technical language rather than overly abstract phrasing.
• Ensure that each portion of the disclosure contributes meaningfully to the search objective.

Key Takeaways

• More descriptive input generally produces better AI search results.
• Context is essential for guiding the system toward the correct technological area.
• Rephrasing, defining terms, describing use, and providing examples all improve performance.
• Background should support the invention—not distract from it.

Well-optimized disclosures enable Quartet AI to more effectively identify relevant documents, improving both recall and precision across the search pipeline.

AutoSearch Cheat Sheet: Best Practices for Drafting Features

The second half of this page outlines how to construct effective features for Quartet Auto Search. Well-structured features improve document screening, scoring, and final relevance ranking across the full retrieval pipeline. We will first briefly discuss how Quartet Auto Search evaluates and analyses documents, then give Auto Search-specific tips on optimizing search inputs.

How Quartet Auto Search Evaluates Documents

It can be helpful to understand how documents are chosen by Quartet Auto Search to better understand how both the full disclosure input and the features can impact the provided results.

Multi-Stage Screening Pipeline

Auto Search evaluates documents through a multi-stage pipeline. In the earliest stage, clearly irrelevant documents are removed—specifically those that show weak alignment across all features. This early filtering reduces noise and allows the system to focus on stronger candidates. As documents move forward, they are examined more closely, with increasingly detailed analysis applied at each step. By the final stages, only higher-quality candidates remain, which improves overall precision.

• Documents with consistently weak matches across all features are removed during initial filtering to reduce noise.
• Remaining documents undergo increasingly detailed analysis, ensuring computational effort is focused on promising candidates.

Independent Feature Matching and Scoring

Each feature is evaluated independently rather than as part of a combined requirement. Documents are scored based on how well they match each feature, and those scores are then considered together. The system tends to prioritize documents that perform well across all features, but it also preserves documents that show particularly strong relevance to one or two features. This balance helps capture both broadly relevant disclosures and highly specific ones that might otherwise be missed. This dual-path retention helps preserve both broad matches and highly specific disclosures.

• Features are matched separately against document text.
• Documents receive scores based on aggregate performance across features.
• Documents with strong performance across all features are prioritized.
• Documents that exhibit exceptional relevance to one or two features may also be retained.

Layered Matching Techniques

Different matching methods are used as documents move through the pipeline. Early on, semantic matching is used to identify conceptual similarities, allowing for flexible phrasing and synonym recognition. In the middle stage, a specialized reranker (Qwen) refines the ordering of results, helping distinguish between moderately relevant candidates. In the final stage, a large language model evaluates the quality and context of each match, producing the most precise relevance judgments.

• Early-stage semantic matching captures conceptual similarity between features and text and allows flexible phrasing and synonym recognition.
• Mid-stage reranking refines ordering using a specialized ranking model and improves discrimination between moderately relevant candidates.
• For the final stage, a large language model assesses the quality and context of matches and produces the most precise relevance judgments.

Granularity of Feature Matching

Features are typically matched against small segments of text, often just one or two sentences, rather than entire documents. This makes precise, well-constructed features especially important, since they need to align with how information is actually expressed in short spans of text.

• Matching occurs on small text units, typically one or two sentences.
• Precision at the sentence level significantly impacts retrieval quality.

Best Practices for Drafting Features

Use Simple, Single-Concept Features

Quartet Auto Search performs best when features focus on a single idea. Combining multiple elements into one feature can reduce the likelihood that a match will be found within a small text segment. Breaking complex descriptions into separate, simple features makes it much easier for the system to identify relevant material and improves overall recall.

Features perform best when they describe one core idea:

Less effective

• “A device with a camera, a laser, a radar and an oscilloscope”

More effective

• Break into separate features:
• • “A device comprising a camera”
  • “The device includes a laser”
  • “The device includes a radar”
  • “The device includes an oscilloscope”

Why this works

• Matching is performed on small text segments.
• Complex, multi-element features reduce the likelihood that all elements appear together within the same sentence span.
• Simpler features increase recall and matching flexibility.

Provide Immediate Context for Each Feature

Providing clear context within each feature is equally important. Features should indicate what the component belongs to or how it is used, rather than describing it in isolation. For example, tying a mechanism to a specific device or component gives the system a clearer signal and leads to more accurate matching and scoring. Since each feature is evaluated independently, this kind of contextual grounding significantly improves results.

Too vague

• “A slidable mechanism”

Improved:

• “The camera includes a slidable mechanism”

Why this works

• Features are evaluated independently.
• Context anchors the feature to a specific component or system.
• Improves scoring accuracy during reranking and LLM evaluation.

Additional Drafting Tips

In general, features should be written in clear, natural technical language. Overly abstract and functional claim language phrasing, such as “means for”, tends to reduce semantic clarity and make matching less effective. Concrete, descriptive language that resembles typical patent disclosures yields better performance. It is also helpful to ensure that each feature adds distinct value, rather than repeating the same idea in slightly different ways.

• Avoid functional claim language such as “means for”
• • Such phrasing is often abstract and reduces semantic clarity.
  • Concrete structural descriptions yield better matches.
• Favor natural technical language
• • Reflect how disclosures are typically written in patent specifications.
  • Avoid overly formal or legalistic constructions when possible.
• Ensure feature distinctiveness
• • Each feature should contribute unique retrieval value.
  • Redundant variations dilute scoring effectiveness.

Summary

Overall, strong Auto Search results depend on clear, focused, and well-contextualized features that align with sentence-level matching and the system’s multi-stage evaluation process. When features are written with these principles in mind, the system is much more effective at identifying relevant disclosures and ranking them appropriately.

• Use simple, focused features
• Include clear contextual grounding
• Align with sentence-level matching behavior
• Leverage understanding of the multi-stage evaluation pipeline

Following these practices leads to stronger matches, better ranking, and more reliable identification of relevant disclosures.