Process results with AI
Last updated
Last updated
Within the Truffle Hunt Clauses mode, the AutoSuggest functionality and your Quality Library, you can filter down your clauses using the help of Artificial Intelligence:
How this works is mostly self-explanatory, but let's go through some of the details nevertheless.
Essentially ClauseBuddy will send your currently visible clauses to a Large Language Model (LLM, aka the new type of "GenAI" Artificial Intelligence), along with your search query or selected text. The AI will then be asked to judge each of those clauses, filter them where necessary, reorder them and send back the ordering.
Truffle Hunt, AutoSuggest and the Quality Library use "smart" search by default, meaning that a superficial layer of artificial intelligence gets to look at your clauses to provide some level of intelligence that goes beyond a literal ("dumb") matching, as is the case with the "strict search". This level of intelligence is very fast (can process thousand results in less than a second) but the downside that the level of intelligence is superficial.
When you involve GenAI, however, a very deep level of intelligence is looking at your clauses. This takes several seconds and cannot encompass hundreds of clauses, but the results are much better though-out.
So all in all it's a balancing act with the current state of the technology, between strict search (super fast but dumb & literal), smart search (the default: reasonably fast, some intelligence) and post-processing with GenAI (slow, but very intelligent). As with all things in life, there are tradeoffs.
The Process results with AI button starts from your current selection of clauses.
The AI post-processing is currently limited to the top 50 clauses. While this amount may go up in the future (when GenAI gets faster) it is therefore very important that you first do a pre-selection with the other tools at your availability, such as the length & content filters in Truffle Hunt and the attributes in the Quality Library.
Once you have a rough selection, you then click on one of the options.
ClauseBuddy will send your search query along with the truffles/clauses to the GenAI-engine. When the query contains multiple words, this is usually enough for the GenAI to figure our what you mean — for example, when your search query would be "liability invoice maximum", then the GenAI will probably figure out itself that you're looking for clauses that relate to liability and some invoicing maximum.
However, you can help the LLM in its job to filter your clauses by providing an additional hint in the Optional prompt box. For example, you can type in there "Only clauses limiting the liability of the supplier", or perhaps something like "don't include clauses relating to payment defaults".
When you click this button, you'll see the GenAI group your clauses into different categories, with each clause in the same category sharing some legal "features". Clauses can belong to multiple categories at once.
For example, in the screenshot below, you can see that about fifteen categories of clauses were created — such as "insurance requirements" and "procurement procedures".
You can then select one or more categories of legal features to filter down what's visible on your screen. For example, if you select "Information security", you'll notice that only three clauses are selected, that all contain something related to information security.
If you then click another (non-light) category, you would further filter down. For example, if you also click on "Data Access and Processing Restrictions", you'll see only those clauses that both talk about information security and data access/processing restrictions.
A light-coloured category means "with your current selection of categories, you would end up with no results", because there are no clauses that belong to all those categories at once.
You can click on the small X at the right to remove the legal features view and return to the normal view.
AutoSuggest is all about getting clauses that are somewhat similar to your currently selected text. The most important post-processing functionality is therefore to simply reorder your current results, so that the most relevant ones — i.e. the most similar ones — are shown first.
The LLM will essentially take your currently selected text, and then figure out how much the clauses in the visible results relate to that selected text.
The Keep if option allows you to write down the criterion for keeping clauses in your result.
If, for example, your confidentiality clauses were extracted from both employment contracts and shareholder agreements, then you could envisage writing "about employees & employers" if you only want to keep the employment-related clauses.
This is the opposite of Keep if. If most results are OK but you want to get rid of a few, go for Remove if.
This button (located in the dropdown button below the brain-icon ) will filter your search results, so that only the clauses/truffles are kept that meet the filter.
This button (located in the dropdown button below the brain-icon ) will compare your currently visible clauses and allow you to easily make subselections.
