At IPRally, our focus has been, and continues to be, on building the most accurate and explainable patent AI search engine. The aim is to automate labor-intensive search tasks from search queries to insightful analyses and reports. 

This is a challenging endeavor, done in phases using technologies best suited for each task. While our own graph search technology is taking leaps to capture all relevant documents, a review of the results is still needed to narrow down to the most relevant hits. 

To address this need and more, we recently introduced a feature called Multi-patent Ask AI, which takes advantage of Large Language Models (LLMs) capability to process complex information quickly and reliably. It works in parallel for dozens or even hundreds of patents, giving you unprecedented power: never before has it been possible to review search results, references in office actions or oppositions, or entire patent portfolios with such flexibility and efficiency.

Below, I introduce ten different ways to use this feature, from summarization to detailed extraction and ad hoc classification. I’ve included some video clips to help you see how smoothly it works in practice. If you prefer to watch the full video briefly covering all the use cases, you can do so here.

Please note that everything you see in the video is achievable for any patent dataset, in any field of technology, and with much more complex questions and technical details. In the video, I use a simple graph search related to liquid containers with rotatable lids, and relatively straightforward questions. 

These are just for inspiration. Ask whatever you need to know. Get creative. 

‍

1. Targeted summarization of patents

Let’s start with an easy, but very helpful use case. Patent abstracts are known to everyone, but they seldom capture the particular interesting points in specific cases. This is solved by Ask AI, which allows you to focus the summary on any particular technical aspect. For example:

• Summarize lid rotation and position-related aspects of the patent.
• Summarize hinge assembly-related aspects.
• What does the patent talk about manufacturing of the device?
• Summarize the independent claims in layman's terms.
• Summarize the method claims in max 20 words.

You get the idea…

2. Multi-aspect summarization: problem and solution

Quite often, one needs to build an overall perception of the key contents of patents, e.g. the problem it solves and proposed main solutions to the problem. Then, one can go with a prompt like this:

1. Which problem does the invention solve?

2. How does it solve the problem?

3. What are the benefits?

4. Summarize the independent claims.

‍

The output is a very digestible overview of the patent. Handy for R&D engineers, perhaps. 

3. Multi-aspect summarization: embodiments

Embodiments described in a patent often provide a clear picture of the patent's scope and relevance. But they are usually buried in lengthy descriptions, and difficult to comprehend. By requesting a summary of the invention and the main embodiments, along with answers to specific critical questions, you get precisely the information you need.

1. What is the key invention of the patent and its benefits?
2. What are the main embodiments and their benefits?
3. Does any of the embodiments address water tightness?

Chemists, just replace “embodiments” with “examples”, and there you go.

4. Extraction of details: The presence of a feature

Often in patent searching, the mere presence of a feature is interesting, either as such, or as an indicator of whether to delve deeper into the contents of a patent. However, finding even a simple feature with Boolean methods often requires defining synonyms, wildcards, etc. This is particularly true for functional features, like whether something is movable, removable or static, as there are multiple ways of expressing such things, and sometimes the disclosure is implicit. Ask AI helps in those challenges by using its built-in understanding of technology and common sense. For example:

• Does the patent disclose a removable lid?
• Does the lid contain an aperture for drinking?
• Is there a liquid surface level indicator?

5. Extraction of details: Interaction of features

Interplay between features is even harder to search and review with traditional methods, without reading the patents thoroughly. Ask AI provides a fast solution: 

1. How is the locking between the lid and the body implemented in the closed position? Are there any pre-tensioned springs? 
2. Does the patent discuss the lid opening automatically and remaining in the open position after its release? 
3. Which axis is the lid rotatable about in relation to the body of the device?

6. Extraction of details: Functional features

Searching for functionally defined features and identifying them during the review phase is cumbersome with traditional methods. Functionally defined features are prevalent across all fields of technology, particularly in software, electronics, and telecommunications where they are practically a must. 

Ask AI makes it possible, and even straightforward, to find these features because the AI understands the meaning of disclosures, not just their literal words. Actually, the user does not need to think about it much. Any of the earlier examples captures functional definitions too. However, let’s take a look at more explicit, yet simple examples, which illustrate type of review questions that are completely out of the scope of traditional review methods, such as keyword highlighting:

• How is the temperature sensor functionally connected to other parts of the device?
• Are there any functionalities in the device that depend on the temperature of the liquid?
• Is the measured temperature transmitted to any external devices? How?

7. Extraction of details: numerical ranges

Chemists, in particular, know that searching for specific numerical values, ranges and range overlaps is hard with Boolean methods. There are two difficult aspects: the number of(range) matching as such, and the use of different units. Even simple Kelvin, Celsius and Fahrenheit units and temperature ranges often cause grey hair.

Numerical values, ranges, and logic (not to mention mathematics involving these) have traditionally posed challenges for AI as well. However, modern Large Language Models demonstrate a relatively strong capability in understanding these complexities for patent review purposes.

• Are any temperature values or ranges disclosed?
• Does the patent discuss temperatures above 95 C? 
• Are any near-infrared wavelengths used? 
• Does any of the examples of the patent show a viscosity of 1.5 mPa·s or more at room temperature?
• What glass transition temperatures are disclosed?

8. Extraction of details: secondary aspects and weak signals

Often, patents contain a lot of secondary or speculative information, that is not in the core of the disclosure, but spotting it may help a lot for example when building argumentation for an opposition case. These aspects can easily remain unnoticed, as they often are sidetracks in the disclosure. Ask AI can assist in detecting those aspects and weak signals. Let’s try two questions, relating to safety and energy saving.

• Does the patent discuss or hint towards protecting the safety of the user of the device? How? 
• Does the patent discuss or hint towards saving energy? How?

The video below exemplifies this well, in particular with the second question, where there are no explicit disclosures. Most of the patents are silent about energy saving as such, but contain some related aspects, that suggest the authors have considered energy consumption. Those can be important links in an inventive step argumentation, for example. 

9. Building structured claim summaries

Ask AI is capable of building structured outputs, too, and pretty good at following meta-level instructions (e.g. relating to bulleting, indentation or answer lengths). One example of this capability is its ability to generate a concise and easily understandable summary of patent claims:

• Build an indented claim tree in this format: [Claim X]: (summary of the claim in max six words). Indent dependent claims.
• List the main manufacturing steps in max five bullets. 

Whether such questions are useful or not, is for you to decide. But at least it’s possible.

10. On-demand patent classification

Sometimes you need to quickly categorize patents into groups to make sense of search results or large data sets. Ask AI is pretty powerful in that. A simple ad hoc example that I use in the video:
‍

Which of the following category or categories the patent falls into (if any)? Answer briefly, with the relevant category label or labels only, followed by short reasoning in [brackets].

A) Nursing device

B) Food warming device

C) Drink warming device

D) Food storage device

Or if the category names are not descriptive enough to make the distinction:
‍

Which of the following category or categories the patent falls into (if any)? Answer briefly, with the relevant category label or labels only, followed by short reasoning in [brackets].

A) Nursing device (designed particularly for infants)

B) Food warming device (heating of foodstuff is mentioned)

C) Drink warming device (heating of water or other liquids is mentioned)

D) Food storage device (long-term storage of foodstuff in the container is explicitly mentioned)

‍

Just replace the categories, and optional descriptions, with what makes sense to you. Note: we have not extensively tested this kind of zero-shot classification, but the initial feeling is positive. Some customers of ours have also been investigating this, with cautiously positive initial reactions. Give it a test!

(Note: For regular classification needs, using established taxonomies and historical training data, IPRally offers a separate AI Classifier module.)

More features coming up

Hopefully the examples above spark some thoughts and pave the way for patent review of the future. As always, we are open for feedback and ideas. One of the things that is already cooking, is smart filtering of the results based on the answers. More on that later.

If you don’t have an access yet to IPRally to try it out, our team is happy to help you. 

If you are a user but have not enabled the Ask AI feature, you, or your organization’s admin user, can do so in the Settings panel.  

Happy asking!

‍