🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.
TypeScript analyticsautogenevaluationlangchainlarge-language-modelsllama-indexllmllm-evaluationllm-observabilityllmopsmonitoringobservabilityopen-sourceopenaiplaygroundprompt-engineeringprompt-managementself-hostedycombinator
import { getLLMErrorInfo } from "../llm/errors";
import { embedLLMText, type LLMProviderOptions } from "../llm/llmText";
import { LLMAdapter } from "../llm/types";
import { topicEmbeddingConfigSchema } from "../../topics";
import type { TopicsModel } from "./model-config";
/** Embeds summaries through the project's LLM connection in as few requests as the provider allows. */
export async function generateTopicEmbeddings(params: {
model: TopicsModel;
summaries: string[];
dimensions: number;
}) {
const { embeddingDimensions } = topicEmbeddingConfigSchema.parse({
embeddingDimensions: params.dimensions,
});
const result = await embedWithSizeFallback(params, embeddingDimensions);
return { embeddings: result.embeddings, tokens: result.usage.tokens };
}
async function embedWithSizeFallback(
params: Parameters<typeof generateTopicEmbeddings>[0],
embeddingDimensions: number,
) {
const embed = (dimensions: number | undefined) =>
embedLLMText({
model: { adapter: params.model.adapter, id: params.model.model },
connection: params.model.connection,
values: params.summaries,
dimensions,
providerOptions: embeddingProviderOptions(params.model),
maxRetries: 2,
// Single-text models (Bedrock Titan, Nova) split a batch into one request per text.
maxParallelCalls: 10,
abortSignal: AbortSignal.timeout(60_000),
});
try {
return await embed(embeddingDimensions);
} catch (error) {
const info = getLLMErrorInfo(error);
if (info?.kind === "provider" && info.statusCode === 400)
return embed(undefined);
throw error;
}
}
/** Keeps discovery and centroid assignment in the same clustering vector space. */
function embeddingProviderOptions(model: TopicsModel): LLMProviderOptions {
if (model.adapter === LLMAdapter.Bedrock)
return { amazonBedrock: bedrockEmbeddingOptions(model.model) };
const google = {
taskType: "CLUSTERING",
};
return {
google,
vertex: google,
};
}
function bedrockEmbeddingOptions(modelId: string): Record<string, string> {
if (modelId.includes("nova"))
return {
embeddingPurpose: "CLUSTERING",
};
if (modelId.includes("cohere"))
return {
inputType: "clustering",
truncate: "NONE",
};
return {};
}