Tools

Tools are built-in capabilities the platform exposes to chat completions. The model can invoke them like any other tool, and the platform handles the execution, the formatting of the results, and the audit trail.

Tools vs. Firewalls

Tools are invoked by the model during the completion (tool calling). Firewalls are a separate mechanism: they are applied before the completion, can rewrite or block the prompt, and never produce a tool call. Both are available on every chat backend.

Compatibility

The same set of tools is available on every chat backend. Pick the backend that fits the workload and enable the tools you need at request time.

Backend	Tools available
Compute (chat instances)	web search, open, find, vector database search
Public Models	web search, open, find, vector database search
Smart Balancer routes	web search, open, find, vector database search

The execution happens inside the inference plane, so the model does not need to know which backend it is running on. The tool definitions look identical to the model regardless of source.

Available tools

The current set of tools. Each one is a flat descriptor on the tools array of a chat completion; the type field tells the platform which tool it is. The browser, search, and find tools are re-exposed to the model as standard function tools during the completion; the vector database tool is executed by the platform before the model is called.

Web search

search

Searches the public web and returns the formatted results. The platform injects a function tool definition with `query` (required), `topn`, and `source` parameters; the model invokes it during the completion and the platform executes the search.

Field	Required	Type	Detail
(none)	no		Just pass {"type":"search"}. The platform injects the function signature for the model.

Open

open

Opens a URL or scrolls within the most recently opened page. Passed flat to the API and re-exposed to the model as a function tool with the same fields.

Field	Required	Type	Detail
id	no	string \| number	Link id from a result page, or a fully qualified URL
cursor	no	number	Page reference; defaults to the most recent page
loc	no	number	Line number to position the viewport
num_lines	no	number	Number of lines to show
view_source	no	boolean	Render the page source instead of the rendered HTML

Find

find

Finds exact matches of a pattern in the current page. Passed flat to the API and re-exposed to the model as a function tool.

Field	Required	Type	Detail
pattern	no	string	Text to find in the page
cursor	no	number	Page reference; defaults to the most recent page

Vector database search

flexible_vector_database_search

Semantic search over a private Vector Database. The platform executes the search itself before the model is called and feeds the hits back as context — the model never sees the tool call.

Field	Required	Type	Detail
vector_database	yes	string	Slug of the database
query	yes	string	Search query
top_k	no	number	Number of hits to return

Request shape

Tools go in the standard tools field of a chat completion, alongside any user-defined function tools you want to register. The type field tells the platform which kind of tool each entry is.

request.json

1{
2  "model": "public:deepseek-v3.2",
3  "messages": [
4    { "role": "user", "content": "What's our refund policy?" }
5  ],
6  "tools": [
7    { "type": "search" },
8    {
9      "type": "flexible_vector_database_search",
10      "vector_database": "support-kb",
11      "query": "refund policy",
12      "top_k": 5
13    },
14    {
15      "type": "function",
16      "function": {
17        "name": "open_ticket",
18        "description": "Open a support ticket in the CRM",
19        "parameters": {
20          "type": "object",
21          "properties": {
22            "subject": { "type": "string" },
23            "body": { "type": "string" }
24          },
25          "required": ["subject"]
26        }
27      }
28    }
29  ]
30}

A tool whose type matches a built-in (one of search, open, find, flexible_vector_database_search) is handled by the platform. A tool with type: "function" is a user-defined tool that is passed through to the model unchanged. You can mix the two in the same request.

Limitations

Chat workloads only: tools are available on chat completions. Embeddings and other workloads do not expose them.
No tool result persistence: the platform executes each tool call inline; results are not cached across requests.
One vector database per call: eachflexible_vector_database_search tool targets a single database. To search across multiple databases, register one tool descriptor per database in the same tools array.