Tools

A model on its own generates text from what it was given. Tool calling lets it look up information and take actions beyond what it was given: open a file, search a folder, render a sheet, query a BIM model.

Think of advising a colleague on site by phone. You can only go on what you already know. If you ask them to send a photo of the condition, the detail number, and the spec paragraph, you can give a better answer. Tool calling is the model making that request, and the photo it gets back becomes part of its context.

How it works​

The application defines a set of tools. Each has a name, a description of when to use it, and the inputs it needs. When the model needs something it does not have, it writes a structured request, such as Read File, this document, pages 4 to 6. The application runs it, the result goes into the context, and the model continues from there.

Any chat assistant that "searches the web" or "runs code" is making a tool call.

The Nomic agent's tools​

The Nomic agent uses a core set of tools on nearly every task. From the agent guide:

Tool	What it does
Search Files	Finds relevant files and passages across the project using Nomic embeddings. Matches by meaning, not exact words.
Read File	Reads pages, sheets, or a whole document through Nomic Parse.
Render Page	Puts the page pixels in front of a vision model for geometry, symbols, and callouts the text layer misses.
List Directory	Explores folders to find the right file.
Artifacts	Creates and edits structured outputs (documents, tables, drawing reviews) beside the conversation.

Read File and Render Page differ in what the model gets back. Read File returns what Nomic Parse extracted from the PDF: structured text, tables, and figures, including OCR of scanned pages. Render Page shows the page itself to a vision model, so it can read geometry and symbols.

Beyond these, the agent runs exact-text and regular-expression searches across attached files. Exact-text search is how it pins a phrase before reading around it, and how it cites a line rather than a page.

Specialized work adds more tools: locating geometry on a sheet, placing markups, verifying a finding with the drawing judge, querying an IFC model or DWG, running code in a sandbox, dispatching sub-agents. Codes Search, which retrieves adopted code text for a jurisdiction, is a tool you turn on from Integrations/Tools on the input bar. Each tool call appears as a step in the conversation as the agent works.

What a tool call costs​

The definitions of the available tools take up space in the context window on every prompt. Every result a tool returns is added to the context window as well. A rendered page adds many tokens, and a full spec division adds more.

Heavy tool use fills the context window faster and costs more. The alternative is a model answering from memory, so the trade-off is nearly always worth it. Tool use is also why a review that reads fifty sheets costs more than a question that reads one.

What tools let you ask​

Without tools, the model is limited to what you paste in:

• "Here is the room finish schedule as text. Which rooms get resilient tile?"

With tools, it can cross-reference:

• "Which rooms on Level 2 have a finish in the schedule that the spec sections do not cover?"

The second question requires finding the schedule, finding the spec sections, reading both, and comparing. Without tools, you would have to paste all of it in yourself.

Next: an agent calling many tools in sequence, deciding what to do from each result.

Next: Agents and sub-agents