Reaching a tool

The assistant has 175 tools and can call any of them by name. This page is about which ones it is told about before it answers, which is a question of speed and cost, never of permission.

These are the shipped defaults. What you can change lives in a config file you own once TESSERACT is installed, so your copy may differ, and should, once you tune it.

Every tool is callable. Only some are described first

Calling a tool by name works whether or not the assistant was told about it at the start of the turn. What differs is how it gets there: a tool it was already shown, it uses right away. A tool it was not shown, it has to find first, which costs one extra step on some models and nothing at all on others.

The working set is the dial you turn

working_set.yaml is where you keep the tools you want described in full on every turn, complete with their inputs. Right now it holds 39 of the 175 tools the assistant has. Anything on it is always at hand.

Move a tool onto the list and it becomes instant to reach, at the cost of a slightly larger request on every turn from then on. Move one off and requests get a little smaller, at the cost of one extra step the first time a conversation needs it. The choice is yours to make and change, and nothing else about what the assistant may do changes with it.

Everything else is grouped by what it is for

The other 136 tools are sorted into 27 groups, each named for the kind of question it answers: remembering something, searching the web, showing you something on screen, and so on. The assistant can reach into any group and use anything inside it without asking your permission first. Being in a group only changes how it gets there, never whether it is allowed to.

It gets there differently depending on which model is answering

Some models see only the working set and the name of each group. To use a tool from outside the working set, they have to say which one they want, wait for it to become available, and then call it. That is one extra step.

Other models see more inside the very same reply: the working set in full, and beside it, every group together with a short list of what is inside. When one of those tools turns out to be needed, the model can bring it in and use it in that same reply, without the extra step, because the group already said enough for it to ask for the right one by name.

Neither way changes what the assistant is allowed to do. What a tool may actually do is decided separately, on What it asks before doing, and stays the same whichever model is answering.

What it costs

Tool descriptions are a real share of what a turn costs before you have typed a word. A tool in the working set costs about 393 tokens of every request, because its full description and its inputs travel along whether or not it gets used that turn. A tool outside the working set costs a great deal less, because it is priced through its group’s short description rather than through a full definition of its own, and a group holding many tools costs little more than one holding few. On the roster this release ships: every tool described in full comes to 59,327 tokens a request, and the same roster with most of them priced through their groups comes to 16,713.

That is the trade the working set makes visible. A tool you reach for constantly earns its keep on every turn; one you rarely touch is cheaper left off the list and found when it is actually needed. What a turn sends opens one request up and shows where the rest of the tokens go.