LibreChat/packages/api/src/agents/run.ts
Danny Avila c87422a1e0
🧠 feat: Thinking Budget, Include Thoughts, and Dynamic Thinking for Gemini 2.5 (#8055)
* feat: support thinking budget parameter for Gemini 2.5 series (#6949, #7542)

https://ai.google.dev/gemini-api/docs/thinking#set-budget

* refactor: update thinking budget minimum value to -1 for dynamic thinking

- see: https://ai.google.dev/gemini-api/docs/thinking#set-budget

* chore: bump @librechat/agents to v2.4.43

* refactor: rename LLMConfigOptions to OpenAIConfigOptions for clarity and consistency

- Updated type definitions and references in initialize.ts, llm.ts, and openai.ts to reflect the new naming convention.
- Ensured that the OpenAI configuration options are consistently used across the relevant files.

* refactor: port Google LLM methods to TypeScript Package

* chore: update @librechat/agents version to 2.4.43 in package-lock.json and package.json

* refactor: update thinking budget description for clarity and adjust placeholder in parameter settings

* refactor: enhance googleSettings default value for thinking budget to support dynamic adjustment

* chore: update @librechat/agents to v2.4.44 for Vertex Dynamic Thinking workaround

* refactor: rename google config function, update `createRun` types, use `reasoning` as `reasoningKey` for Google

* refactor: simplify placeholder handling in DynamicInput component

* refactor: enhance thinking budget description for clarity and allow automatic decision by setting to "-1"

* refactor: update text styling in OptionHover component for improved readability

* chore: update @librechat/agents dependency to v2.4.46 in package.json and package-lock.json

* chore: update @librechat/api version to 1.2.5 in package.json and package-lock.json

* refactor: enhance `clientOptions` handling by filtering `omitTitleOptions`, add `json` field for Google models

---------

Co-authored-by: ciffelia <15273128+ciffelia@users.noreply.github.com>
2025-06-25 15:14:33 -04:00

101 lines
2.8 KiB
TypeScript

import { Run, Providers } from '@librechat/agents';
import { providerEndpointMap, KnownEndpoints } from 'librechat-data-provider';
import type {
StandardGraphConfig,
EventHandler,
GenericTool,
GraphEvents,
IState,
} from '@librechat/agents';
import type { Agent } from 'librechat-data-provider';
import type * as t from '~/types';
const customProviders = new Set([
Providers.XAI,
Providers.OLLAMA,
Providers.DEEPSEEK,
Providers.OPENROUTER,
]);
/**
* Creates a new Run instance with custom handlers and configuration.
*
* @param options - The options for creating the Run instance.
* @param options.agent - The agent for this run.
* @param options.signal - The signal for this run.
* @param options.req - The server request.
* @param options.runId - Optional run ID; otherwise, a new run ID will be generated.
* @param options.customHandlers - Custom event handlers.
* @param options.streaming - Whether to use streaming.
* @param options.streamUsage - Whether to stream usage information.
* @returns {Promise<Run<IState>>} A promise that resolves to a new Run instance.
*/
export async function createRun({
runId,
agent,
signal,
customHandlers,
streaming = true,
streamUsage = true,
}: {
agent: Omit<Agent, 'tools'> & { tools?: GenericTool[] };
signal: AbortSignal;
runId?: string;
streaming?: boolean;
streamUsage?: boolean;
customHandlers?: Record<GraphEvents, EventHandler>;
}): Promise<Run<IState>> {
const provider =
(providerEndpointMap[
agent.provider as keyof typeof providerEndpointMap
] as unknown as Providers) ?? agent.provider;
const llmConfig: t.RunLLMConfig = Object.assign(
{
provider,
streaming,
streamUsage,
},
agent.model_parameters,
);
/** Resolves issues with new OpenAI usage field */
if (
customProviders.has(agent.provider) ||
(agent.provider === Providers.OPENAI && agent.endpoint !== agent.provider)
) {
llmConfig.streamUsage = false;
llmConfig.usage = true;
}
let reasoningKey: 'reasoning_content' | 'reasoning' | undefined;
if (provider === Providers.GOOGLE) {
reasoningKey = 'reasoning';
} else if (
llmConfig.configuration?.baseURL?.includes(KnownEndpoints.openrouter) ||
(agent.endpoint && agent.endpoint.toLowerCase().includes(KnownEndpoints.openrouter))
) {
reasoningKey = 'reasoning';
}
const graphConfig: StandardGraphConfig = {
signal,
llmConfig,
reasoningKey,
tools: agent.tools,
instructions: agent.instructions,
additional_instructions: agent.additional_instructions,
// toolEnd: agent.end_after_tools,
};
// TEMPORARY FOR TESTING
if (agent.provider === Providers.ANTHROPIC || agent.provider === Providers.BEDROCK) {
graphConfig.streamBuffer = 2000;
}
return Run.create({
runId,
graphConfig,
customHandlers,
});
}