@dbx-tools/shared-model

Validate model lookups and ranked model responses in UIs, routes, tools, and tests without importing Databricks clients. Use the browser-safe model selection, capability, reasoning, and OpenAI Chat Completions contracts across client and server boundaries. Server-side normalization, classification, ranking, and fuzzy resolution live in @dbx-tools/model.

Format model display names

import { display } from "@dbx-tools/shared-model";

display.toModelDisplayName("databricks-claude-sonnet-4-6"); // "Claude Sonnet 4.6"
display.toModelDisplayName("system.ai.bge_large_en"); // "BGE Large En"
display.toModelDisplayName("x", "Claude 4.6 (Preview)"); // provided name wins

ServingEndpointSummary.displayName is the optional friendly label for the picker. name stays the invocation ID. A Databricks-provided name wins. Otherwise, the pure helper strips leading vendor prefixes and title-cases through @dbx-tools/shared-core's tokenizer.

The package also includes generated ModelClass, endpoint, query, status, profile, ranking, and reasoning-effort contracts, plus browser-safe OpenAI Chat Completions message and tool-call contracts.

Validate a model lookup request

import { model } from "@dbx-tools/shared-model";

const query = model.ModelQuerySchema.parse({
  search: "claude sonnet",
  modelClass: "chat-balanced",
  requiresTools: true,
  limit: 5,
});

Use model.ModelQuerySchema for route query/body validation and agent tool inputs. It keeps client model pickers and backend resolution endpoints on the same request shape.

Use classified endpoint catalogs

import { model } from "@dbx-tools/shared-model";

const fast = endpoints.filter((endpoint) => endpoint.class === model.ModelClass.ChatFast);
const agentModels = endpoints.filter((endpoint) => endpoint.supportsTools);

The server stamps class and capability fields before returning /models. Browser clients consume those fields directly. Model-family policy does not run in the browser.

Strip fields that Databricks rejects

import { openaiChat } from "@dbx-tools/shared-model";

const dropped = openaiChat.stripUnsupportedChatFields(body);

Databricks Model Serving validates Chat Completions bodies strictly. stripUnsupportedChatFields deletes known unsupported fields in place and returns the removed names so a caller can log the difference.

Runtime boundary

This package doesn't discover endpoints, read credentials, rank models, or send Model Serving requests. Use @dbx-tools/model for server-side model discovery and selection. Use @dbx-tools/shared-model-gateway for browser-safe model gateway responses and discovery clients.

Package exports

• model: model classes, reasoning-effort schema, endpoint summaries, lookup requests, and ranked results.
• display: human-readable endpoint labels.
• openaiChat: Chat Completions message and tool-call types, content helpers, and rejected-field sanitization.