diff --git a/.changeset/early-crabs-nail.md b/.changeset/early-crabs-nail.md
new file mode 100644
index 0000000000000000000000000000000000000000..7978fc549d5a1c52d89b9537c470bf1f5620d113
--- /dev/null
+++ b/.changeset/early-crabs-nail.md
@@ -0,0 +1,8 @@
+---
+"llamaindex": patch
+"@llamaindex/openai": patch
+---
+
+feat: decouple openai from `llamaindex` module
+
+This should be a non-breaking change, but just you can now only install `@llamaindex/openai` to reduce the bundle size in the future
diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 6325a1a6017ed6c05a3cf8b6e8a9e29e1849bb96..2e48f5e4b8e98042ebadf8c5b9c982d2cae4435c 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -142,6 +142,9 @@ jobs:
- name: Pack @llamaindex/cloud
run: pnpm pack --pack-destination ${{ runner.temp }}
working-directory: packages/cloud
+ - name: Pack @llamaindex/openai
+ run: pnpm pack --pack-destination ${{ runner.temp }}
+ working-directory: packages/llm/openai
- name: Pack @llamaindex/core
run: pnpm pack --pack-destination ${{ runner.temp }}
working-directory: packages/core
diff --git a/packages/llamaindex/e2e/fixtures/embeddings/OpenAIEmbedding.ts b/packages/llamaindex/e2e/fixtures/embeddings/OpenAIEmbedding.ts
deleted file mode 100644
index 2ff111e2e05b4532a924c5184f558557df6c33f7..0000000000000000000000000000000000000000
--- a/packages/llamaindex/e2e/fixtures/embeddings/OpenAIEmbedding.ts
+++ /dev/null
@@ -1,51 +0,0 @@
-import { TransformComponent } from "@llamaindex/core/schema";
-import {
- BaseEmbedding,
- BaseNode,
- SimilarityType,
- type EmbeddingInfo,
- type MessageContentDetail,
-} from "llamaindex";
-
-export class OpenAIEmbedding
- extends TransformComponent
- implements BaseEmbedding
-{
- embedInfo?: EmbeddingInfo;
- embedBatchSize = 512;
-
- constructor() {
- super(async (nodes: BaseNode[], _options?: any): Promise<BaseNode[]> => {
- nodes.forEach((node) => (node.embedding = [0]));
- return nodes;
- });
- }
-
- async getQueryEmbedding(query: MessageContentDetail) {
- return [0];
- }
-
- async getTextEmbedding(text: string) {
- return [0];
- }
-
- async getTextEmbeddings(texts: string[]) {
- return [[0]];
- }
-
- async getTextEmbeddingsBatch(texts: string[]) {
- return [[0]];
- }
-
- similarity(
- embedding1: number[],
- embedding2: number[],
- mode?: SimilarityType,
- ) {
- return 1;
- }
-
- truncateMaxTokens(input: string[]): string[] {
- return input;
- }
-}
diff --git a/packages/llamaindex/e2e/fixtures/llm/openai.ts b/packages/llamaindex/e2e/fixtures/llm/openai.ts
index 2b54919f5e5ad7f70762770886a3f4d7ead2fd66..10fd90c6453dc9ccd5dbe84ba932923bf2e4ddaf 100644
--- a/packages/llamaindex/e2e/fixtures/llm/openai.ts
+++ b/packages/llamaindex/e2e/fixtures/llm/openai.ts
@@ -12,6 +12,15 @@ import type {
import { deepStrictEqual, strictEqual } from "node:assert";
import { llmCompleteMockStorage } from "../../node/utils.js";
+import { TransformComponent } from "@llamaindex/core/schema";
+import {
+ BaseEmbedding,
+ BaseNode,
+ SimilarityType,
+ type EmbeddingInfo,
+ type MessageContentDetail,
+} from "llamaindex";
+
export function getOpenAISession() {
return {};
}
@@ -22,6 +31,7 @@ export function isFunctionCallingModel() {
export class OpenAI implements LLM {
supportToolCall = true;
+
get metadata() {
return {
model: "mock-model",
@@ -32,6 +42,7 @@ export class OpenAI implements LLM {
isFunctionCallingModel: true,
};
}
+
chat(
params: LLMChatParamsStreaming<Record<string, unknown>>,
): Promise<AsyncIterable<ChatResponseChunk>>;
@@ -77,6 +88,7 @@ export class OpenAI implements LLM {
}
throw new Error("Method not implemented.");
}
+
complete(
params: LLMCompletionParamsStreaming,
): Promise<AsyncIterable<CompletionResponse>>;
@@ -103,3 +115,46 @@ export class OpenAI implements LLM {
throw new Error("Method not implemented.");
}
}
+
+export class OpenAIEmbedding
+ extends TransformComponent
+ implements BaseEmbedding
+{
+ embedInfo?: EmbeddingInfo;
+ embedBatchSize = 512;
+
+ constructor() {
+ super(async (nodes: BaseNode[], _options?: any): Promise<BaseNode[]> => {
+ nodes.forEach((node) => (node.embedding = [0]));
+ return nodes;
+ });
+ }
+
+ async getQueryEmbedding(query: MessageContentDetail) {
+ return [0];
+ }
+
+ async getTextEmbedding(text: string) {
+ return [0];
+ }
+
+ async getTextEmbeddings(texts: string[]) {
+ return [[0]];
+ }
+
+ async getTextEmbeddingsBatch(texts: string[]) {
+ return [[0]];
+ }
+
+ similarity(
+ embedding1: number[],
+ embedding2: number[],
+ mode?: SimilarityType,
+ ) {
+ return 1;
+ }
+
+ truncateMaxTokens(input: string[]): string[] {
+ return input;
+ }
+}
diff --git a/packages/llamaindex/e2e/mock-module.js b/packages/llamaindex/e2e/mock-module.js
index d32e8f1878f9aca65f8ca8d3171a36ddbc3d4f0a..fe6cd62629c8c21d1f601df8d0ec295db16a44f2 100644
--- a/packages/llamaindex/e2e/mock-module.js
+++ b/packages/llamaindex/e2e/mock-module.js
@@ -13,8 +13,14 @@ export async function resolve(specifier, context, nextResolve) {
return result;
}
const targetUrl = fileURLToPath(result.url).replace(/\.js$/, ".ts");
- const relativePath = relative(packageDistDir, targetUrl);
- if (relativePath.startsWith(".") || relativePath.startsWith("/")) {
+ let relativePath = relative(packageDistDir, targetUrl);
+ // todo: make it more generic if we have more sub modules fixtures in the future
+ if (relativePath.startsWith("../../llm/openai")) {
+ relativePath = relativePath.replace(
+ "../../llm/openai/dist/index.ts",
+ "llm/openai.ts",
+ );
+ } else if (relativePath.startsWith(".") || relativePath.startsWith("/")) {
return result;
}
const url = pathToFileURL(join(fixturesDir, relativePath)).toString();
diff --git a/packages/llamaindex/e2e/package.json b/packages/llamaindex/e2e/package.json
index 971b369f1c85b46b3bc41022f63ddd05d467d9e9..0f3f588ea45521bd8e684ae50c9be139beca83df 100644
--- a/packages/llamaindex/e2e/package.json
+++ b/packages/llamaindex/e2e/package.json
@@ -10,7 +10,6 @@
},
"devDependencies": {
"@faker-js/faker": "^8.4.1",
- "@llamaindex/core": "workspace:*",
"@types/node": "^22.5.1",
"consola": "^3.2.3",
"llamaindex": "workspace:*",
diff --git a/packages/llamaindex/package.json b/packages/llamaindex/package.json
index 711907fd19b03e5491c4a89dc14460bdf3ffb3ea..436ded68465bc0af64d34de78de7f5a7a2b89ffb 100644
--- a/packages/llamaindex/package.json
+++ b/packages/llamaindex/package.json
@@ -33,6 +33,7 @@
"@llamaindex/cloud": "workspace:*",
"@llamaindex/core": "workspace:*",
"@llamaindex/env": "workspace:*",
+ "@llamaindex/openai": "workspace:*",
"@mistralai/mistralai": "^1.0.4",
"@mixedbread-ai/sdk": "^2.2.11",
"@pinecone-database/pinecone": "^3.0.2",
diff --git a/packages/llamaindex/src/ChatHistory.ts b/packages/llamaindex/src/ChatHistory.ts
index f681e5261eb88762305a8ec78c9c65d1c26d86f7..4bbd44e4b491a85da240b6c91bcb1876cdaa0d40 100644
--- a/packages/llamaindex/src/ChatHistory.ts
+++ b/packages/llamaindex/src/ChatHistory.ts
@@ -5,7 +5,7 @@ import {
} from "@llamaindex/core/prompts";
import { extractText, messagesToHistory } from "@llamaindex/core/utils";
import { tokenizers, type Tokenizer } from "@llamaindex/env";
-import { OpenAI } from "./llm/openai.js";
+import { OpenAI } from "@llamaindex/openai";
/**
* A ChatHistory is used to keep the state of back and forth chat messages
diff --git a/packages/llamaindex/src/QuestionGenerator.ts b/packages/llamaindex/src/QuestionGenerator.ts
index 47ea6001eea4dc4dc1ae0582845723c2fe3b49d7..0f69fdfb843a562f346c9ddd40a5dd938efbec0c 100644
--- a/packages/llamaindex/src/QuestionGenerator.ts
+++ b/packages/llamaindex/src/QuestionGenerator.ts
@@ -8,12 +8,12 @@ import {
import type { QueryType } from "@llamaindex/core/query-engine";
import type { BaseOutputParser } from "@llamaindex/core/schema";
import { extractText, toToolDescriptions } from "@llamaindex/core/utils";
+import { OpenAI } from "@llamaindex/openai";
import { SubQuestionOutputParser } from "./OutputParser.js";
import type {
BaseQuestionGenerator,
SubQuestion,
} from "./engines/query/types.js";
-import { OpenAI } from "./llm/openai.js";
import type { StructuredOutput } from "./types.js";
/**
diff --git a/packages/llamaindex/src/ServiceContext.ts b/packages/llamaindex/src/ServiceContext.ts
index c6c1885f295dc3af95b3e4046b5a1b7319c4a4a9..a185423fbdcb68759c9049dbea9514322f82fd89 100644
--- a/packages/llamaindex/src/ServiceContext.ts
+++ b/packages/llamaindex/src/ServiceContext.ts
@@ -5,8 +5,7 @@ import {
type NodeParser,
SentenceSplitter,
} from "@llamaindex/core/node-parser";
-import { OpenAIEmbedding } from "./embeddings/OpenAIEmbedding.js";
-import { OpenAI } from "./llm/openai.js";
+import { OpenAI, OpenAIEmbedding } from "@llamaindex/openai";
/**
* The ServiceContext is a collection of components that are used in different parts of the application.
diff --git a/packages/llamaindex/src/Settings.ts b/packages/llamaindex/src/Settings.ts
index a4acce25d429fbb24a5af24d3b1daa2e4bb1f946..a39fc3b1ab05c427b6e264df68b793556eef8e1d 100644
--- a/packages/llamaindex/src/Settings.ts
+++ b/packages/llamaindex/src/Settings.ts
@@ -2,7 +2,7 @@ import {
type CallbackManager,
Settings as CoreSettings,
} from "@llamaindex/core/global";
-import { OpenAI } from "./llm/openai.js";
+import { OpenAI } from "@llamaindex/openai";
import { PromptHelper } from "@llamaindex/core/indices";
diff --git a/packages/llamaindex/src/agent/openai.ts b/packages/llamaindex/src/agent/openai.ts
index a85fb4c5a1719984dddaaca8cdf0d66c55897c5b..8cd84bc675ff3972df36eec3b91d6fa158686f31 100644
--- a/packages/llamaindex/src/agent/openai.ts
+++ b/packages/llamaindex/src/agent/openai.ts
@@ -1,5 +1,5 @@
+import { OpenAI } from "@llamaindex/openai";
import { Settings } from "../Settings.js";
-import { OpenAI } from "../llm/openai.js";
import { LLMAgent, LLMAgentWorker, type LLMAgentParams } from "./llm.js";
// This is likely not necessary anymore but leaving it here just incase it's in use elsewhere
diff --git a/packages/llamaindex/src/cloud/LlamaCloudIndex.ts b/packages/llamaindex/src/cloud/LlamaCloudIndex.ts
index 7e83ab1360501cf170eea56aa2143cfd7cb6ecd3..14ddb6f8de3884807b08d50f7d48fb4c1f2b8f95 100644
--- a/packages/llamaindex/src/cloud/LlamaCloudIndex.ts
+++ b/packages/llamaindex/src/cloud/LlamaCloudIndex.ts
@@ -13,8 +13,8 @@ import { getAppBaseUrl, getProjectId, initService } from "./utils.js";
import { PipelinesService, ProjectsService } from "@llamaindex/cloud/api";
import { SentenceSplitter } from "@llamaindex/core/node-parser";
import { getEnv } from "@llamaindex/env";
+import { OpenAIEmbedding } from "@llamaindex/openai";
import { Settings } from "../Settings.js";
-import { OpenAIEmbedding } from "../embeddings/OpenAIEmbedding.js";
export class LlamaCloudIndex {
params: CloudConstructorParams;
diff --git a/packages/llamaindex/src/cloud/config.ts b/packages/llamaindex/src/cloud/config.ts
index 0a562fd58662c94c584d4f3c6fd56a2c6d2ef2c8..5b3b7a48e38f101943fbc22a98d8c0e037b7d575 100644
--- a/packages/llamaindex/src/cloud/config.ts
+++ b/packages/llamaindex/src/cloud/config.ts
@@ -5,7 +5,7 @@ import type {
} from "@llamaindex/cloud/api";
import { SentenceSplitter } from "@llamaindex/core/node-parser";
import { BaseNode, type TransformComponent } from "@llamaindex/core/schema";
-import { OpenAIEmbedding } from "../embeddings/OpenAIEmbedding.js";
+import { OpenAIEmbedding } from "@llamaindex/openai";
export type GetPipelineCreateParams = {
pipelineName: string;
diff --git a/packages/llamaindex/src/embeddings/OpenAIEmbedding.ts b/packages/llamaindex/src/embeddings/OpenAIEmbedding.ts
index 72386410e1dfd3a97568bf31c6f709ffcaaa0e55..02781efcb42bde6e5598eb279af193c85ec9158f 100644
--- a/packages/llamaindex/src/embeddings/OpenAIEmbedding.ts
+++ b/packages/llamaindex/src/embeddings/OpenAIEmbedding.ts
@@ -1,152 +1 @@
-import { BaseEmbedding } from "@llamaindex/core/embeddings";
-import { Tokenizers } from "@llamaindex/env";
-import type { ClientOptions as OpenAIClientOptions } from "openai";
-import type { AzureOpenAIConfig } from "../llm/azure.js";
-import {
- getAzureConfigFromEnv,
- getAzureModel,
- shouldUseAzure,
-} from "../llm/azure.js";
-import type { OpenAISession } from "../llm/openai.js";
-import { getOpenAISession } from "../llm/openai.js";
-
-export const ALL_OPENAI_EMBEDDING_MODELS = {
- "text-embedding-ada-002": {
- dimensions: 1536,
- maxTokens: 8192,
- tokenizer: Tokenizers.CL100K_BASE,
- },
- "text-embedding-3-small": {
- dimensions: 1536,
- dimensionOptions: [512, 1536],
- maxTokens: 8192,
- tokenizer: Tokenizers.CL100K_BASE,
- },
- "text-embedding-3-large": {
- dimensions: 3072,
- dimensionOptions: [256, 1024, 3072],
- maxTokens: 8192,
- tokenizer: Tokenizers.CL100K_BASE,
- },
-};
-
-type ModelKeys = keyof typeof ALL_OPENAI_EMBEDDING_MODELS;
-
-export class OpenAIEmbedding extends BaseEmbedding {
- /** embeddding model. defaults to "text-embedding-ada-002" */
- model: string;
- /** number of dimensions of the resulting vector, for models that support choosing fewer dimensions. undefined will default to model default */
- dimensions?: number | undefined;
-
- // OpenAI session params
-
- /** api key */
- apiKey?: string | undefined = undefined;
- /** maximum number of retries, default 10 */
- maxRetries: number;
- /** timeout in ms, default 60 seconds */
- timeout?: number | undefined;
- /** other session options for OpenAI */
- additionalSessionOptions?:
- | Omit<Partial<OpenAIClientOptions>, "apiKey" | "maxRetries" | "timeout">
- | undefined;
-
- /** session object */
- session: OpenAISession;
-
- /**
- * OpenAI Embedding
- * @param init - initial parameters
- */
- constructor(init?: Partial<OpenAIEmbedding> & { azure?: AzureOpenAIConfig }) {
- super();
-
- this.model = init?.model ?? "text-embedding-ada-002";
- this.dimensions = init?.dimensions; // if no dimensions provided, will be undefined/not sent to OpenAI
-
- this.embedBatchSize = init?.embedBatchSize ?? 10;
- this.maxRetries = init?.maxRetries ?? 10;
-
- this.timeout = init?.timeout ?? 60 * 1000; // Default is 60 seconds
- this.additionalSessionOptions = init?.additionalSessionOptions;
-
- // find metadata for model
- const key = Object.keys(ALL_OPENAI_EMBEDDING_MODELS).find(
- (key) => key === this.model,
- ) as ModelKeys | undefined;
- if (key) {
- this.embedInfo = ALL_OPENAI_EMBEDDING_MODELS[key];
- }
-
- if (init?.azure || shouldUseAzure()) {
- const azureConfig = {
- ...getAzureConfigFromEnv({
- model: getAzureModel(this.model),
- }),
- ...init?.azure,
- };
-
- this.apiKey = azureConfig.apiKey;
- this.session =
- init?.session ??
- getOpenAISession({
- azure: true,
- maxRetries: this.maxRetries,
- timeout: this.timeout,
- ...this.additionalSessionOptions,
- ...azureConfig,
- });
- } else {
- this.apiKey = init?.apiKey ?? undefined;
- this.session =
- init?.session ??
- getOpenAISession({
- apiKey: this.apiKey,
- maxRetries: this.maxRetries,
- timeout: this.timeout,
- ...this.additionalSessionOptions,
- });
- }
- }
-
- /**
- * Get embeddings for a batch of texts
- * @param texts
- * @param options
- */
- private async getOpenAIEmbedding(input: string[]): Promise<number[][]> {
- // TODO: ensure this for every sub class by calling it in the base class
- input = this.truncateMaxTokens(input);
-
- const { data } = await this.session.openai.embeddings.create(
- this.dimensions
- ? {
- model: this.model,
- dimensions: this.dimensions, // only sent to OpenAI if set by user
- input,
- }
- : {
- model: this.model,
- input,
- },
- );
-
- return data.map((d) => d.embedding);
- }
-
- /**
- * Get embeddings for a batch of texts
- * @param texts
- */
- getTextEmbeddings = async (texts: string[]): Promise<number[][]> => {
- return this.getOpenAIEmbedding(texts);
- };
-
- /**
- * Get embeddings for a single text
- * @param texts
- */
- async getTextEmbedding(text: string): Promise<number[]> {
- return (await this.getOpenAIEmbedding([text]))[0]!;
- }
-}
+export * from "@llamaindex/openai";
diff --git a/packages/llamaindex/src/embeddings/fireworks.ts b/packages/llamaindex/src/embeddings/fireworks.ts
index a48039e1022e91c06561e08937a02b39bc3f5e7e..8338884c12955937b009a28c1d482d6799a2bfd7 100644
--- a/packages/llamaindex/src/embeddings/fireworks.ts
+++ b/packages/llamaindex/src/embeddings/fireworks.ts
@@ -1,5 +1,5 @@
import { getEnv } from "@llamaindex/env";
-import { OpenAIEmbedding } from "./OpenAIEmbedding.js";
+import { OpenAIEmbedding } from "@llamaindex/openai";
export class FireworksEmbedding extends OpenAIEmbedding {
constructor(init?: Partial<OpenAIEmbedding>) {
diff --git a/packages/llamaindex/src/embeddings/together.ts b/packages/llamaindex/src/embeddings/together.ts
index b284daf6a701fe42000b3184dfc68c5f62d7e127..1ed43fef7f582bb557c9afe961fec41eb1c182b0 100644
--- a/packages/llamaindex/src/embeddings/together.ts
+++ b/packages/llamaindex/src/embeddings/together.ts
@@ -1,5 +1,5 @@
import { getEnv } from "@llamaindex/env";
-import { OpenAIEmbedding } from "./OpenAIEmbedding.js";
+import { OpenAIEmbedding } from "@llamaindex/openai";
export class TogetherEmbedding extends OpenAIEmbedding {
constructor(init?: Partial<OpenAIEmbedding>) {
diff --git a/packages/llamaindex/src/extractors/MetadataExtractors.ts b/packages/llamaindex/src/extractors/MetadataExtractors.ts
index b7ca6c49f9c8fef157ef172b2c22ddaa6134d828..b1d0ce5a47412627745c0375ef86b1a0f30e27b9 100644
--- a/packages/llamaindex/src/extractors/MetadataExtractors.ts
+++ b/packages/llamaindex/src/extractors/MetadataExtractors.ts
@@ -1,7 +1,7 @@
import type { LLM } from "@llamaindex/core/llms";
import type { BaseNode } from "@llamaindex/core/schema";
import { MetadataMode, TextNode } from "@llamaindex/core/schema";
-import { OpenAI } from "../llm/index.js";
+import { OpenAI } from "@llamaindex/openai";
import {
defaultKeywordExtractorPromptTemplate,
defaultQuestionAnswerPromptTemplate,
diff --git a/packages/llamaindex/src/internal/settings/EmbedModel.ts b/packages/llamaindex/src/internal/settings/EmbedModel.ts
index fab2331a8702e0d077330258a5c18d404e2f743e..d912d20b9c4121891751f1317d347b4ca9ca9dfc 100644
--- a/packages/llamaindex/src/internal/settings/EmbedModel.ts
+++ b/packages/llamaindex/src/internal/settings/EmbedModel.ts
@@ -1,6 +1,6 @@
import type { BaseEmbedding } from "@llamaindex/core/embeddings";
import { AsyncLocalStorage } from "@llamaindex/env";
-import { OpenAIEmbedding } from "../../embeddings/OpenAIEmbedding.js";
+import { OpenAIEmbedding } from "@llamaindex/openai";
const embeddedModelAsyncLocalStorage = new AsyncLocalStorage<BaseEmbedding>();
let globalEmbeddedModel: BaseEmbedding | null = null;
diff --git a/packages/llamaindex/src/llm/deepinfra.ts b/packages/llamaindex/src/llm/deepinfra.ts
index 9e2d9e2f8b20958552c07b23d907cc256f112d41..c2c8bde814bd8ef1be6ae0a39abb6d2e53c3edca 100644
--- a/packages/llamaindex/src/llm/deepinfra.ts
+++ b/packages/llamaindex/src/llm/deepinfra.ts
@@ -1,5 +1,5 @@
import { getEnv } from "@llamaindex/env";
-import { OpenAI } from "./openai.js";
+import { OpenAI } from "@llamaindex/openai";
const ENV_VARIABLE_NAME = "DEEPINFRA_API_TOKEN";
const DEFAULT_MODEL = "mistralai/Mixtral-8x22B-Instruct-v0.1";
diff --git a/packages/llamaindex/src/llm/deepseek.ts b/packages/llamaindex/src/llm/deepseek.ts
index 7c4f15466df2d81192197a61a96b33dcdc1d10b2..d8a2586838ab763e1a0f98a07f9746f274b407d7 100644
--- a/packages/llamaindex/src/llm/deepseek.ts
+++ b/packages/llamaindex/src/llm/deepseek.ts
@@ -1,5 +1,5 @@
import { getEnv } from "@llamaindex/env";
-import { OpenAI } from "./openai.js";
+import { OpenAI } from "@llamaindex/openai";
export const DEEPSEEK_MODELS = {
"deepseek-coder": { contextWindow: 128000 },
diff --git a/packages/llamaindex/src/llm/fireworks.ts b/packages/llamaindex/src/llm/fireworks.ts
index 0385b62f25206b8eb2fa3d2cdcdef0115abbfd8d..3e5979f4d921708c01016ef4a24f190e8b7e5a84 100644
--- a/packages/llamaindex/src/llm/fireworks.ts
+++ b/packages/llamaindex/src/llm/fireworks.ts
@@ -1,5 +1,5 @@
import { getEnv } from "@llamaindex/env";
-import { OpenAI } from "./openai.js";
+import { OpenAI } from "@llamaindex/openai";
export class FireworksLLM extends OpenAI {
constructor(init?: Partial<OpenAI>) {
diff --git a/packages/llamaindex/src/llm/groq.ts b/packages/llamaindex/src/llm/groq.ts
index 8f77a9389cebcc24c04c2e996f779b73fb6c88b0..5c058f923ed7847cadbf02654fc4a3c3ebea3a7b 100644
--- a/packages/llamaindex/src/llm/groq.ts
+++ b/packages/llamaindex/src/llm/groq.ts
@@ -1,6 +1,6 @@
import { getEnv } from "@llamaindex/env";
+import { OpenAI } from "@llamaindex/openai";
import GroqSDK, { type ClientOptions } from "groq-sdk";
-import { OpenAI } from "./openai.js";
export class Groq extends OpenAI {
constructor(
diff --git a/packages/llamaindex/src/llm/index.ts b/packages/llamaindex/src/llm/index.ts
index de4dbad211b0b58c4979935c6f76cb6bca3c76a9..722a006657d0c8672a79aebd01e4b56de9dd7014 100644
--- a/packages/llamaindex/src/llm/index.ts
+++ b/packages/llamaindex/src/llm/index.ts
@@ -11,7 +11,6 @@ export {
GEMINI_MODEL,
type GoogleGeminiSessionOptions,
} from "./gemini/types.js";
-
export { Groq } from "./groq.js";
export { HuggingFaceInferenceAPI, HuggingFaceLLM } from "./huggingface.js";
export {
diff --git a/packages/llamaindex/src/llm/openai.ts b/packages/llamaindex/src/llm/openai.ts
index 1a5b237a160d346465a05402611e79381e92cd2f..02781efcb42bde6e5598eb279af193c85ec9158f 100644
--- a/packages/llamaindex/src/llm/openai.ts
+++ b/packages/llamaindex/src/llm/openai.ts
@@ -1,502 +1 @@
-import { getEnv } from "@llamaindex/env";
-import _ from "lodash";
-import type OpenAILLM from "openai";
-import type {
- ClientOptions,
- ClientOptions as OpenAIClientOptions,
-} from "openai";
-import { AzureOpenAI, OpenAI as OrigOpenAI } from "openai";
-import type { ChatModel } from "openai/resources/chat/chat";
-
-import {
- type BaseTool,
- type ChatMessage,
- type ChatResponse,
- type ChatResponseChunk,
- type LLM,
- type LLMChatParamsNonStreaming,
- type LLMChatParamsStreaming,
- type LLMMetadata,
- type MessageType,
- type PartialToolCall,
- ToolCallLLM,
- type ToolCallLLMMessageOptions,
-} from "@llamaindex/core/llms";
-import {
- extractText,
- wrapEventCaller,
- wrapLLMEvent,
-} from "@llamaindex/core/utils";
-import { Tokenizers } from "@llamaindex/env";
-import type {
- ChatCompletionAssistantMessageParam,
- ChatCompletionMessageToolCall,
- ChatCompletionRole,
- ChatCompletionSystemMessageParam,
- ChatCompletionTool,
- ChatCompletionToolMessageParam,
- ChatCompletionUserMessageParam,
-} from "openai/resources/chat/completions";
-import type { ChatCompletionMessageParam } from "openai/resources/index.js";
-import type { AzureOpenAIConfig } from "./azure.js";
-import {
- getAzureConfigFromEnv,
- getAzureModel,
- shouldUseAzure,
-} from "./azure.js";
-
-export class OpenAISession {
- openai: Pick<OrigOpenAI, "chat" | "embeddings">;
-
- constructor(options: ClientOptions & { azure?: boolean } = {}) {
- if (options.azure) {
- this.openai = new AzureOpenAI(options as AzureOpenAIConfig);
- } else {
- if (!options.apiKey) {
- options.apiKey = getEnv("OPENAI_API_KEY");
- }
-
- if (!options.apiKey) {
- throw new Error("Set OpenAI Key in OPENAI_API_KEY env variable"); // Overriding OpenAI package's error message
- }
-
- this.openai = new OrigOpenAI({
- ...options,
- });
- }
- }
-}
-
-// I'm not 100% sure this is necessary vs. just starting a new session
-// every time we make a call. They say they try to reuse connections
-// so in theory this is more efficient, but we should test it in the future.
-const defaultOpenAISession: {
- session: OpenAISession;
- options: ClientOptions;
-}[] = [];
-
-/**
- * Get a session for the OpenAI API. If one already exists with the same options,
- * it will be returned. Otherwise, a new session will be created.
- * @param options
- * @returns
- */
-export function getOpenAISession(
- options: ClientOptions & { azure?: boolean } = {},
-) {
- let session = defaultOpenAISession.find((session) => {
- return _.isEqual(session.options, options);
- })?.session;
-
- if (!session) {
- session = new OpenAISession(options);
- defaultOpenAISession.push({ session, options });
- }
-
- return session;
-}
-
-export const GPT4_MODELS = {
- "chatgpt-4o-latest": {
- contextWindow: 128000,
- },
- "gpt-4": { contextWindow: 8192 },
- "gpt-4-32k": { contextWindow: 32768 },
- "gpt-4-32k-0613": { contextWindow: 32768 },
- "gpt-4-turbo": { contextWindow: 128000 },
- "gpt-4-turbo-preview": { contextWindow: 128000 },
- "gpt-4-1106-preview": { contextWindow: 128000 },
- "gpt-4-0125-preview": { contextWindow: 128000 },
- "gpt-4-vision-preview": { contextWindow: 128000 },
- "gpt-4o": { contextWindow: 128000 },
- "gpt-4o-2024-05-13": { contextWindow: 128000 },
- "gpt-4o-mini": { contextWindow: 128000 },
- "gpt-4o-mini-2024-07-18": { contextWindow: 128000 },
- "gpt-4o-2024-08-06": { contextWindow: 128000 },
- "gpt-4o-2024-09-14": { contextWindow: 128000 },
- "gpt-4o-2024-10-14": { contextWindow: 128000 },
- "gpt-4-0613": { contextWindow: 128000 },
- "gpt-4-turbo-2024-04-09": { contextWindow: 128000 },
- "gpt-4-0314": { contextWindow: 128000 },
- "gpt-4-32k-0314": { contextWindow: 32768 },
-};
-
-// NOTE we don't currently support gpt-3.5-turbo-instruct and don't plan to in the near future
-export const GPT35_MODELS = {
- "gpt-3.5-turbo": { contextWindow: 16385 },
- "gpt-3.5-turbo-0613": { contextWindow: 4096 },
- "gpt-3.5-turbo-16k": { contextWindow: 16385 },
- "gpt-3.5-turbo-16k-0613": { contextWindow: 16385 },
- "gpt-3.5-turbo-1106": { contextWindow: 16385 },
- "gpt-3.5-turbo-0125": { contextWindow: 16385 },
- "gpt-3.5-turbo-0301": { contextWindow: 16385 },
-};
-
-export const O1_MODELS = {
- "o1-preview": {
- contextWindow: 128000,
- },
- "o1-preview-2024-09-12": {
- contextWindow: 128000,
- },
- "o1-mini": {
- contextWindow: 128000,
- },
- "o1-mini-2024-09-12": {
- contextWindow: 128000,
- },
-};
-
-/**
- * We currently support GPT-3.5 and GPT-4 models
- */
-export const ALL_AVAILABLE_OPENAI_MODELS = {
- ...GPT4_MODELS,
- ...GPT35_MODELS,
- ...O1_MODELS,
-} satisfies Record<ChatModel, { contextWindow: number }>;
-
-export function isFunctionCallingModel(llm: LLM): llm is OpenAI {
- let model: string;
- if (llm instanceof OpenAI) {
- model = llm.model;
- } else if ("model" in llm && typeof llm.model === "string") {
- model = llm.model;
- } else {
- return false;
- }
- const isChatModel = Object.keys(ALL_AVAILABLE_OPENAI_MODELS).includes(model);
- const isOld = model.includes("0314") || model.includes("0301");
- const isO1 = model.startsWith("o1");
- return isChatModel && !isOld && !isO1;
-}
-
-export type OpenAIAdditionalMetadata = {};
-
-export type OpenAIAdditionalChatOptions = Omit<
- Partial<OpenAILLM.Chat.ChatCompletionCreateParams>,
- | "max_tokens"
- | "messages"
- | "model"
- | "temperature"
- | "top_p"
- | "stream"
- | "tools"
- | "toolChoice"
->;
-
-export class OpenAI extends ToolCallLLM<OpenAIAdditionalChatOptions> {
- model:
- | ChatModel
- // string & {} is a hack to allow any string, but still give autocomplete
- | (string & {});
- temperature: number;
- topP: number;
- maxTokens?: number | undefined;
- additionalChatOptions?: OpenAIAdditionalChatOptions | undefined;
-
- // OpenAI session params
- apiKey?: string | undefined = undefined;
- maxRetries: number;
- timeout?: number;
- session: OpenAISession;
- additionalSessionOptions?:
- | undefined
- | Omit<Partial<OpenAIClientOptions>, "apiKey" | "maxRetries" | "timeout">;
-
- constructor(
- init?: Partial<OpenAI> & {
- azure?: AzureOpenAIConfig;
- },
- ) {
- super();
- this.model = init?.model ?? "gpt-4o";
- this.temperature = init?.temperature ?? 0.1;
- this.topP = init?.topP ?? 1;
- this.maxTokens = init?.maxTokens ?? undefined;
-
- this.maxRetries = init?.maxRetries ?? 10;
- this.timeout = init?.timeout ?? 60 * 1000; // Default is 60 seconds
- this.additionalChatOptions = init?.additionalChatOptions;
- this.additionalSessionOptions = init?.additionalSessionOptions;
-
- if (init?.azure || shouldUseAzure()) {
- const azureConfig = {
- ...getAzureConfigFromEnv({
- model: getAzureModel(this.model),
- }),
- ...init?.azure,
- };
-
- this.apiKey = azureConfig.apiKey;
- this.session =
- init?.session ??
- getOpenAISession({
- azure: true,
- maxRetries: this.maxRetries,
- timeout: this.timeout,
- ...this.additionalSessionOptions,
- ...azureConfig,
- });
- } else {
- this.apiKey = init?.apiKey ?? undefined;
- this.session =
- init?.session ??
- getOpenAISession({
- apiKey: this.apiKey,
- maxRetries: this.maxRetries,
- timeout: this.timeout,
- ...this.additionalSessionOptions,
- });
- }
- }
-
- get supportToolCall() {
- return isFunctionCallingModel(this);
- }
-
- get metadata(): LLMMetadata & OpenAIAdditionalMetadata {
- const contextWindow =
- ALL_AVAILABLE_OPENAI_MODELS[
- this.model as keyof typeof ALL_AVAILABLE_OPENAI_MODELS
- ]?.contextWindow ?? 1024;
- return {
- model: this.model,
- temperature: this.temperature,
- topP: this.topP,
- maxTokens: this.maxTokens,
- contextWindow,
- tokenizer: Tokenizers.CL100K_BASE,
- };
- }
-
- static toOpenAIRole(messageType: MessageType): ChatCompletionRole {
- switch (messageType) {
- case "user":
- return "user";
- case "assistant":
- return "assistant";
- case "system":
- return "system";
- default:
- return "user";
- }
- }
-
- static toOpenAIMessage(
- messages: ChatMessage<ToolCallLLMMessageOptions>[],
- ): ChatCompletionMessageParam[] {
- return messages.map((message) => {
- const options = message.options ?? {};
- if ("toolResult" in options) {
- return {
- tool_call_id: options.toolResult.id,
- role: "tool",
- content: extractText(message.content),
- } satisfies ChatCompletionToolMessageParam;
- } else if ("toolCall" in options) {
- return {
- role: "assistant",
- content: extractText(message.content),
- tool_calls: options.toolCall.map((toolCall) => {
- return {
- id: toolCall.id,
- type: "function",
- function: {
- name: toolCall.name,
- arguments:
- typeof toolCall.input === "string"
- ? toolCall.input
- : JSON.stringify(toolCall.input),
- },
- };
- }),
- } satisfies ChatCompletionAssistantMessageParam;
- } else if (message.role === "user") {
- return {
- role: "user",
- content: message.content,
- } satisfies ChatCompletionUserMessageParam;
- }
-
- const response:
- | ChatCompletionSystemMessageParam
- | ChatCompletionUserMessageParam
- | ChatCompletionMessageToolCall = {
- // fixme(alex): type assertion
- role: OpenAI.toOpenAIRole(message.role) as never,
- // fixme: should not extract text, but assert content is string
- content: extractText(message.content),
- };
- return response;
- });
- }
-
- chat(
- params: LLMChatParamsStreaming<
- OpenAIAdditionalChatOptions,
- ToolCallLLMMessageOptions
- >,
- ): Promise<AsyncIterable<ChatResponseChunk<ToolCallLLMMessageOptions>>>;
- chat(
- params: LLMChatParamsNonStreaming<
- OpenAIAdditionalChatOptions,
- ToolCallLLMMessageOptions
- >,
- ): Promise<ChatResponse<ToolCallLLMMessageOptions>>;
- @wrapEventCaller
- @wrapLLMEvent
- async chat(
- params:
- | LLMChatParamsNonStreaming<
- OpenAIAdditionalChatOptions,
- ToolCallLLMMessageOptions
- >
- | LLMChatParamsStreaming<
- OpenAIAdditionalChatOptions,
- ToolCallLLMMessageOptions
- >,
- ): Promise<
- | ChatResponse<ToolCallLLMMessageOptions>
- | AsyncIterable<ChatResponseChunk<ToolCallLLMMessageOptions>>
- > {
- const { messages, stream, tools, additionalChatOptions } = params;
- const baseRequestParams = <OpenAILLM.Chat.ChatCompletionCreateParams>{
- model: this.model,
- temperature: this.temperature,
- max_tokens: this.maxTokens,
- tools: tools?.map(OpenAI.toTool),
- messages: OpenAI.toOpenAIMessage(messages),
- top_p: this.topP,
- ...Object.assign({}, this.additionalChatOptions, additionalChatOptions),
- };
-
- if (
- Array.isArray(baseRequestParams.tools) &&
- baseRequestParams.tools.length === 0
- ) {
- // remove empty tools array to avoid OpenAI error
- delete baseRequestParams.tools;
- }
-
- // Streaming
- if (stream) {
- return this.streamChat(baseRequestParams);
- }
-
- // Non-streaming
- const response = await this.session.openai.chat.completions.create({
- ...baseRequestParams,
- stream: false,
- });
-
- const content = response.choices[0]!.message?.content ?? "";
-
- return {
- raw: response,
- message: {
- content,
- role: response.choices[0]!.message.role,
- options: response.choices[0]!.message?.tool_calls
- ? {
- toolCall: response.choices[0]!.message.tool_calls.map(
- (toolCall) => ({
- id: toolCall.id,
- name: toolCall.function.name,
- input: toolCall.function.arguments,
- }),
- ),
- }
- : {},
- },
- };
- }
-
- // todo: this wrapper is ugly, refactor it
- @wrapEventCaller
- protected async *streamChat(
- baseRequestParams: OpenAILLM.Chat.ChatCompletionCreateParams,
- ): AsyncIterable<ChatResponseChunk<ToolCallLLMMessageOptions>> {
- const stream: AsyncIterable<OpenAILLM.Chat.ChatCompletionChunk> =
- await this.session.openai.chat.completions.create({
- ...baseRequestParams,
- stream: true,
- });
-
- // TODO: add callback to streamConverter and use streamConverter here
- // this will be used to keep track of the current tool call, make sure input are valid json object.
- let currentToolCall: PartialToolCall | null = null;
- const toolCallMap = new Map<string, PartialToolCall>();
- for await (const part of stream) {
- if (part.choices.length === 0) continue;
- const choice = part.choices[0]!;
- // skip parts that don't have any content
- if (!(choice.delta.content || choice.delta.tool_calls)) continue;
-
- let shouldEmitToolCall: PartialToolCall | null = null;
- if (
- choice.delta.tool_calls?.[0]!.id &&
- currentToolCall &&
- choice.delta.tool_calls?.[0].id !== currentToolCall.id
- ) {
- shouldEmitToolCall = {
- ...currentToolCall,
- input: JSON.parse(currentToolCall.input),
- };
- }
- if (choice.delta.tool_calls?.[0]!.id) {
- currentToolCall = {
- name: choice.delta.tool_calls[0].function!.name!,
- id: choice.delta.tool_calls[0].id,
- input: choice.delta.tool_calls[0].function!.arguments!,
- };
- toolCallMap.set(choice.delta.tool_calls[0].id, currentToolCall);
- } else {
- if (choice.delta.tool_calls?.[0]!.function?.arguments) {
- currentToolCall!.input +=
- choice.delta.tool_calls[0].function.arguments;
- }
- }
-
- const isDone: boolean = choice.finish_reason !== null;
-
- if (isDone && currentToolCall) {
- // for the last one, we need to emit the tool call
- shouldEmitToolCall = {
- ...currentToolCall,
- input: JSON.parse(currentToolCall.input),
- };
- }
-
- yield {
- raw: part,
- options: shouldEmitToolCall
- ? { toolCall: [shouldEmitToolCall] }
- : currentToolCall
- ? {
- toolCall: [currentToolCall],
- }
- : {},
- delta: choice.delta.content ?? "",
- };
- }
- toolCallMap.clear();
- return;
- }
-
- static toTool(tool: BaseTool): ChatCompletionTool {
- return {
- type: "function",
- function: tool.metadata.parameters
- ? {
- name: tool.metadata.name,
- description: tool.metadata.description,
- parameters: tool.metadata.parameters,
- }
- : {
- name: tool.metadata.name,
- description: tool.metadata.description,
- },
- };
- }
-}
+export * from "@llamaindex/openai";
diff --git a/packages/llamaindex/src/llm/together.ts b/packages/llamaindex/src/llm/together.ts
index 3772b2be2bf08cd2b7735a1419d9e09684c1873c..4d314bcc0a84261580dac5502651dffd7d0f48d6 100644
--- a/packages/llamaindex/src/llm/together.ts
+++ b/packages/llamaindex/src/llm/together.ts
@@ -1,5 +1,5 @@
import { getEnv } from "@llamaindex/env";
-import { OpenAI } from "./openai.js";
+import { OpenAI } from "@llamaindex/openai";
export class TogetherLLM extends OpenAI {
constructor(init?: Partial<OpenAI>) {
diff --git a/packages/llm/openai/package.json b/packages/llm/openai/package.json
new file mode 100644
index 0000000000000000000000000000000000000000..5fb001d0066a61176b754398848f0169fad628df
--- /dev/null
+++ b/packages/llm/openai/package.json
@@ -0,0 +1,42 @@
+{
+ "name": "@llamaindex/openai",
+ "description": "OpenAI Adapter for LlamaIndex",
+ "version": "0.1.0",
+ "type": "module",
+ "main": "./dist/index.cjs",
+ "module": "./dist/index.js",
+ "exports": {
+ ".": {
+ "require": {
+ "types": "./dist/index.d.cts",
+ "default": "./dist/index.cjs"
+ },
+ "import": {
+ "types": "./dist/index.d.ts",
+ "default": "./dist/index.js"
+ }
+ }
+ },
+ "files": [
+ "dist"
+ ],
+ "repository": {
+ "type": "git",
+ "url": "https://github.com/run-llama/LlamaIndexTS.git",
+ "directory": "packages/llm/openai"
+ },
+ "private": true,
+ "scripts": {
+ "build": "bunchee",
+ "dev": "bunchee --watch"
+ },
+ "devDependencies": {
+ "bunchee": "5.3.2"
+ },
+ "dependencies": {
+ "@llamaindex/core": "workspace:*",
+ "@llamaindex/env": "workspace:*",
+ "openai": "^4.60.0",
+ "remeda": "^2.12.0"
+ }
+}
diff --git a/packages/llamaindex/src/llm/azure.ts b/packages/llm/openai/src/azure.ts
similarity index 100%
rename from packages/llamaindex/src/llm/azure.ts
rename to packages/llm/openai/src/azure.ts
diff --git a/packages/llm/openai/src/embedding.ts b/packages/llm/openai/src/embedding.ts
new file mode 100644
index 0000000000000000000000000000000000000000..e4b5024488e5cf1ec6a7a554a38211b6432f6d2c
--- /dev/null
+++ b/packages/llm/openai/src/embedding.ts
@@ -0,0 +1,152 @@
+import { BaseEmbedding } from "@llamaindex/core/embeddings";
+import { Tokenizers } from "@llamaindex/env";
+import type { ClientOptions as OpenAIClientOptions } from "openai";
+import type { AzureOpenAIConfig } from "./azure.js";
+import {
+ getAzureConfigFromEnv,
+ getAzureModel,
+ shouldUseAzure,
+} from "./azure.js";
+import type { OpenAISession } from "./llm.js";
+import { getOpenAISession } from "./llm.js";
+
+export const ALL_OPENAI_EMBEDDING_MODELS = {
+ "text-embedding-ada-002": {
+ dimensions: 1536,
+ maxTokens: 8192,
+ tokenizer: Tokenizers.CL100K_BASE,
+ },
+ "text-embedding-3-small": {
+ dimensions: 1536,
+ dimensionOptions: [512, 1536],
+ maxTokens: 8192,
+ tokenizer: Tokenizers.CL100K_BASE,
+ },
+ "text-embedding-3-large": {
+ dimensions: 3072,
+ dimensionOptions: [256, 1024, 3072],
+ maxTokens: 8192,
+ tokenizer: Tokenizers.CL100K_BASE,
+ },
+};
+
+type ModelKeys = keyof typeof ALL_OPENAI_EMBEDDING_MODELS;
+
+export class OpenAIEmbedding extends BaseEmbedding {
+ /** embeddding model. defaults to "text-embedding-ada-002" */
+ model: string;
+ /** number of dimensions of the resulting vector, for models that support choosing fewer dimensions. undefined will default to model default */
+ dimensions?: number | undefined;
+
+ // OpenAI session params
+
+ /** api key */
+ apiKey?: string | undefined = undefined;
+ /** maximum number of retries, default 10 */
+ maxRetries: number;
+ /** timeout in ms, default 60 seconds */
+ timeout?: number | undefined;
+ /** other session options for OpenAI */
+ additionalSessionOptions?:
+ | Omit<Partial<OpenAIClientOptions>, "apiKey" | "maxRetries" | "timeout">
+ | undefined;
+
+ /** session object */
+ session: OpenAISession;
+
+ /**
+ * OpenAI Embedding
+ * @param init - initial parameters
+ */
+ constructor(init?: Partial<OpenAIEmbedding> & { azure?: AzureOpenAIConfig }) {
+ super();
+
+ this.model = init?.model ?? "text-embedding-ada-002";
+ this.dimensions = init?.dimensions; // if no dimensions provided, will be undefined/not sent to OpenAI
+
+ this.embedBatchSize = init?.embedBatchSize ?? 10;
+ this.maxRetries = init?.maxRetries ?? 10;
+
+ this.timeout = init?.timeout ?? 60 * 1000; // Default is 60 seconds
+ this.additionalSessionOptions = init?.additionalSessionOptions;
+
+ // find metadata for model
+ const key = Object.keys(ALL_OPENAI_EMBEDDING_MODELS).find(
+ (key) => key === this.model,
+ ) as ModelKeys | undefined;
+ if (key) {
+ this.embedInfo = ALL_OPENAI_EMBEDDING_MODELS[key];
+ }
+
+ if (init?.azure || shouldUseAzure()) {
+ const azureConfig = {
+ ...getAzureConfigFromEnv({
+ model: getAzureModel(this.model),
+ }),
+ ...init?.azure,
+ };
+
+ this.apiKey = azureConfig.apiKey;
+ this.session =
+ init?.session ??
+ getOpenAISession({
+ azure: true,
+ maxRetries: this.maxRetries,
+ timeout: this.timeout,
+ ...this.additionalSessionOptions,
+ ...azureConfig,
+ });
+ } else {
+ this.apiKey = init?.apiKey ?? undefined;
+ this.session =
+ init?.session ??
+ getOpenAISession({
+ apiKey: this.apiKey,
+ maxRetries: this.maxRetries,
+ timeout: this.timeout,
+ ...this.additionalSessionOptions,
+ });
+ }
+ }
+
+ /**
+ * Get embeddings for a batch of texts
+ * @param texts
+ * @param options
+ */
+ private async getOpenAIEmbedding(input: string[]): Promise<number[][]> {
+ // TODO: ensure this for every sub class by calling it in the base class
+ input = this.truncateMaxTokens(input);
+
+ const { data } = await this.session.openai.embeddings.create(
+ this.dimensions
+ ? {
+ model: this.model,
+ dimensions: this.dimensions, // only sent to OpenAI if set by user
+ input,
+ }
+ : {
+ model: this.model,
+ input,
+ },
+ );
+
+ return data.map((d) => d.embedding);
+ }
+
+ /**
+ * Get embeddings for a batch of texts
+ * @param texts
+ */
+ getTextEmbeddings = async (texts: string[]): Promise<number[][]> => {
+ return this.getOpenAIEmbedding(texts);
+ };
+
+ /**
+ * Get embeddings for a single text
+ * @param texts
+ */
+ async getTextEmbedding(text: string): Promise<number[]> {
+ return (await this.getOpenAIEmbedding([text]))[0]!;
+ }
+}
diff --git a/packages/llm/openai/src/index.ts b/packages/llm/openai/src/index.ts
new file mode 100644
index 0000000000000000000000000000000000000000..f1e663639c19225d50f1142545b3ac92f75488bf
--- /dev/null
+++ b/packages/llm/openai/src/index.ts
@@ -0,0 +1,13 @@
+export { ALL_OPENAI_EMBEDDING_MODELS, OpenAIEmbedding } from "./embedding";
+export {
+ ALL_AVAILABLE_OPENAI_MODELS,
+ GPT35_MODELS,
+ GPT4_MODELS,
+ O1_MODELS,
+ OpenAI,
+ OpenAISession,
+ type OpenAIAdditionalChatOptions,
+ type OpenAIAdditionalMetadata,
+} from "./llm";
+
+export { type AzureOpenAIConfig } from "./azure";
diff --git a/packages/llm/openai/src/llm.ts b/packages/llm/openai/src/llm.ts
new file mode 100644
index 0000000000000000000000000000000000000000..58d7aa1d22d710450ef3b282c94fd959bc883a9a
--- /dev/null
+++ b/packages/llm/openai/src/llm.ts
@@ -0,0 +1,502 @@
+import { getEnv } from "@llamaindex/env";
+import type OpenAILLM from "openai";
+import type {
+ ClientOptions,
+ ClientOptions as OpenAIClientOptions,
+} from "openai";
+import { AzureOpenAI, OpenAI as OrigOpenAI } from "openai";
+import type { ChatModel } from "openai/resources/chat/chat";
+import { isDeepEqual } from "remeda";
+
+import {
+ type BaseTool,
+ type ChatMessage,
+ type ChatResponse,
+ type ChatResponseChunk,
+ type LLM,
+ type LLMChatParamsNonStreaming,
+ type LLMChatParamsStreaming,
+ type LLMMetadata,
+ type MessageType,
+ type PartialToolCall,
+ ToolCallLLM,
+ type ToolCallLLMMessageOptions,
+} from "@llamaindex/core/llms";
+import {
+ extractText,
+ wrapEventCaller,
+ wrapLLMEvent,
+} from "@llamaindex/core/utils";
+import { Tokenizers } from "@llamaindex/env";
+import type {
+ ChatCompletionAssistantMessageParam,
+ ChatCompletionMessageToolCall,
+ ChatCompletionRole,
+ ChatCompletionSystemMessageParam,
+ ChatCompletionTool,
+ ChatCompletionToolMessageParam,
+ ChatCompletionUserMessageParam,
+} from "openai/resources/chat/completions";
+import type { ChatCompletionMessageParam } from "openai/resources/index.js";
+import type { AzureOpenAIConfig } from "./azure.js";
+import {
+ getAzureConfigFromEnv,
+ getAzureModel,
+ shouldUseAzure,
+} from "./azure.js";
+
+export class OpenAISession {
+ openai: Pick<OrigOpenAI, "chat" | "embeddings">;
+
+ constructor(options: ClientOptions & { azure?: boolean } = {}) {
+ if (options.azure) {
+ this.openai = new AzureOpenAI(options as AzureOpenAIConfig);
+ } else {
+ if (!options.apiKey) {
+ options.apiKey = getEnv("OPENAI_API_KEY");
+ }
+
+ if (!options.apiKey) {
+ throw new Error("Set OpenAI Key in OPENAI_API_KEY env variable"); // Overriding OpenAI package's error message
+ }
+
+ this.openai = new OrigOpenAI({
+ ...options,
+ });
+ }
+ }
+}
+
+// I'm not 100% sure this is necessary vs. just starting a new session
+// every time we make a call. They say they try to reuse connections
+// so in theory this is more efficient, but we should test it in the future.
+const defaultOpenAISession: {
+ session: OpenAISession;
+ options: ClientOptions;
+}[] = [];
+
+/**
+ * Get a session for the OpenAI API. If one already exists with the same options,
+ * it will be returned. Otherwise, a new session will be created.
+ * @param options
+ * @returns
+ */
+export function getOpenAISession(
+ options: ClientOptions & { azure?: boolean } = {},
+) {
+ let session = defaultOpenAISession.find((session) => {
+ return isDeepEqual(session.options, options);
+ })?.session;
+
+ if (!session) {
+ session = new OpenAISession(options);
+ defaultOpenAISession.push({ session, options });
+ }
+
+ return session;
+}
+
+export const GPT4_MODELS = {
+ "chatgpt-4o-latest": {
+ contextWindow: 128000,
+ },
+ "gpt-4": { contextWindow: 8192 },
+ "gpt-4-32k": { contextWindow: 32768 },
+ "gpt-4-32k-0613": { contextWindow: 32768 },
+ "gpt-4-turbo": { contextWindow: 128000 },
+ "gpt-4-turbo-preview": { contextWindow: 128000 },
+ "gpt-4-1106-preview": { contextWindow: 128000 },
+ "gpt-4-0125-preview": { contextWindow: 128000 },
+ "gpt-4-vision-preview": { contextWindow: 128000 },
+ "gpt-4o": { contextWindow: 128000 },
+ "gpt-4o-2024-05-13": { contextWindow: 128000 },
+ "gpt-4o-mini": { contextWindow: 128000 },
+ "gpt-4o-mini-2024-07-18": { contextWindow: 128000 },
+ "gpt-4o-2024-08-06": { contextWindow: 128000 },
+ "gpt-4o-2024-09-14": { contextWindow: 128000 },
+ "gpt-4o-2024-10-14": { contextWindow: 128000 },
+ "gpt-4-0613": { contextWindow: 128000 },
+ "gpt-4-turbo-2024-04-09": { contextWindow: 128000 },
+ "gpt-4-0314": { contextWindow: 128000 },
+ "gpt-4-32k-0314": { contextWindow: 32768 },
+};
+
+// NOTE we don't currently support gpt-3.5-turbo-instruct and don't plan to in the near future
+export const GPT35_MODELS = {
+ "gpt-3.5-turbo": { contextWindow: 16385 },
+ "gpt-3.5-turbo-0613": { contextWindow: 4096 },
+ "gpt-3.5-turbo-16k": { contextWindow: 16385 },
+ "gpt-3.5-turbo-16k-0613": { contextWindow: 16385 },
+ "gpt-3.5-turbo-1106": { contextWindow: 16385 },
+ "gpt-3.5-turbo-0125": { contextWindow: 16385 },
+ "gpt-3.5-turbo-0301": { contextWindow: 16385 },
+};
+
+export const O1_MODELS = {
+ "o1-preview": {
+ contextWindow: 128000,
+ },
+ "o1-preview-2024-09-12": {
+ contextWindow: 128000,
+ },
+ "o1-mini": {
+ contextWindow: 128000,
+ },
+ "o1-mini-2024-09-12": {
+ contextWindow: 128000,
+ },
+};
+
+/**
+ * We currently support GPT-3.5 and GPT-4 models
+ */
+export const ALL_AVAILABLE_OPENAI_MODELS = {
+ ...GPT4_MODELS,
+ ...GPT35_MODELS,
+ ...O1_MODELS,
+} satisfies Record<ChatModel, { contextWindow: number }>;
+
+export function isFunctionCallingModel(llm: LLM): llm is OpenAI {
+ let model: string;
+ if (llm instanceof OpenAI) {
+ model = llm.model;
+ } else if ("model" in llm && typeof llm.model === "string") {
+ model = llm.model;
+ } else {
+ return false;
+ }
+ const isChatModel = Object.keys(ALL_AVAILABLE_OPENAI_MODELS).includes(model);
+ const isOld = model.includes("0314") || model.includes("0301");
+ const isO1 = model.startsWith("o1");
+ return isChatModel && !isOld && !isO1;
+}
+
+export type OpenAIAdditionalMetadata = {};
+
+export type OpenAIAdditionalChatOptions = Omit<
+ Partial<OpenAILLM.Chat.ChatCompletionCreateParams>,
+ | "max_tokens"
+ | "messages"
+ | "model"
+ | "temperature"
+ | "top_p"
+ | "stream"
+ | "tools"
+ | "toolChoice"
+>;
+
+export class OpenAI extends ToolCallLLM<OpenAIAdditionalChatOptions> {
+ model:
+ | ChatModel
+ // string & {} is a hack to allow any string, but still give autocomplete
+ | (string & {});
+ temperature: number;
+ topP: number;
+ maxTokens?: number | undefined;
+ additionalChatOptions?: OpenAIAdditionalChatOptions | undefined;
+
+ // OpenAI session params
+ apiKey?: string | undefined = undefined;
+ maxRetries: number;
+ timeout?: number;
+ session: OpenAISession;
+ additionalSessionOptions?:
+ | undefined
+ | Omit<Partial<OpenAIClientOptions>, "apiKey" | "maxRetries" | "timeout">;
+
+ constructor(
+ init?: Partial<OpenAI> & {
+ azure?: AzureOpenAIConfig;
+ },
+ ) {
+ super();
+ this.model = init?.model ?? "gpt-4o";
+ this.temperature = init?.temperature ?? 0.1;
+ this.topP = init?.topP ?? 1;
+ this.maxTokens = init?.maxTokens ?? undefined;
+
+ this.maxRetries = init?.maxRetries ?? 10;
+ this.timeout = init?.timeout ?? 60 * 1000; // Default is 60 seconds
+ this.additionalChatOptions = init?.additionalChatOptions;
+ this.additionalSessionOptions = init?.additionalSessionOptions;
+
+ if (init?.azure || shouldUseAzure()) {
+ const azureConfig = {
+ ...getAzureConfigFromEnv({
+ model: getAzureModel(this.model),
+ }),
+ ...init?.azure,
+ };
+
+ this.apiKey = azureConfig.apiKey;
+ this.session =
+ init?.session ??
+ getOpenAISession({
+ azure: true,
+ maxRetries: this.maxRetries,
+ timeout: this.timeout,
+ ...this.additionalSessionOptions,
+ ...azureConfig,
+ });
+ } else {
+ this.apiKey = init?.apiKey ?? undefined;
+ this.session =
+ init?.session ??
+ getOpenAISession({
+ apiKey: this.apiKey,
+ maxRetries: this.maxRetries,
+ timeout: this.timeout,
+ ...this.additionalSessionOptions,
+ });
+ }
+ }
+
+ get supportToolCall() {
+ return isFunctionCallingModel(this);
+ }
+
+ get metadata(): LLMMetadata & OpenAIAdditionalMetadata {
+ const contextWindow =
+ ALL_AVAILABLE_OPENAI_MODELS[
+ this.model as keyof typeof ALL_AVAILABLE_OPENAI_MODELS
+ ]?.contextWindow ?? 1024;
+ return {
+ model: this.model,
+ temperature: this.temperature,
+ topP: this.topP,
+ maxTokens: this.maxTokens,
+ contextWindow,
+ tokenizer: Tokenizers.CL100K_BASE,
+ };
+ }
+
+ static toOpenAIRole(messageType: MessageType): ChatCompletionRole {
+ switch (messageType) {
+ case "user":
+ return "user";
+ case "assistant":
+ return "assistant";
+ case "system":
+ return "system";
+ default:
+ return "user";
+ }
+ }
+
+ static toOpenAIMessage(
+ messages: ChatMessage<ToolCallLLMMessageOptions>[],
+ ): ChatCompletionMessageParam[] {
+ return messages.map((message) => {
+ const options = message.options ?? {};
+ if ("toolResult" in options) {
+ return {
+ tool_call_id: options.toolResult.id,
+ role: "tool",
+ content: extractText(message.content),
+ } satisfies ChatCompletionToolMessageParam;
+ } else if ("toolCall" in options) {
+ return {
+ role: "assistant",
+ content: extractText(message.content),
+ tool_calls: options.toolCall.map((toolCall) => {
+ return {
+ id: toolCall.id,
+ type: "function",
+ function: {
+ name: toolCall.name,
+ arguments:
+ typeof toolCall.input === "string"
+ ? toolCall.input
+ : JSON.stringify(toolCall.input),
+ },
+ };
+ }),
+ } satisfies ChatCompletionAssistantMessageParam;
+ } else if (message.role === "user") {
+ return {
+ role: "user",
+ content: message.content,
+ } satisfies ChatCompletionUserMessageParam;
+ }
+
+ const response:
+ | ChatCompletionSystemMessageParam
+ | ChatCompletionUserMessageParam
+ | ChatCompletionMessageToolCall = {
+ // fixme(alex): type assertion
+ role: OpenAI.toOpenAIRole(message.role) as never,
+ // fixme: should not extract text, but assert content is string
+ content: extractText(message.content),
+ };
+ return response;
+ });
+ }
+
+ chat(
+ params: LLMChatParamsStreaming<
+ OpenAIAdditionalChatOptions,
+ ToolCallLLMMessageOptions
+ >,
+ ): Promise<AsyncIterable<ChatResponseChunk<ToolCallLLMMessageOptions>>>;
+ chat(
+ params: LLMChatParamsNonStreaming<
+ OpenAIAdditionalChatOptions,
+ ToolCallLLMMessageOptions
+ >,
+ ): Promise<ChatResponse<ToolCallLLMMessageOptions>>;
+ @wrapEventCaller
+ @wrapLLMEvent
+ async chat(
+ params:
+ | LLMChatParamsNonStreaming<
+ OpenAIAdditionalChatOptions,
+ ToolCallLLMMessageOptions
+ >
+ | LLMChatParamsStreaming<
+ OpenAIAdditionalChatOptions,
+ ToolCallLLMMessageOptions
+ >,
+ ): Promise<
+ | ChatResponse<ToolCallLLMMessageOptions>
+ | AsyncIterable<ChatResponseChunk<ToolCallLLMMessageOptions>>
+ > {
+ const { messages, stream, tools, additionalChatOptions } = params;
+ const baseRequestParams = <OpenAILLM.Chat.ChatCompletionCreateParams>{
+ model: this.model,
+ temperature: this.temperature,
+ max_tokens: this.maxTokens,
+ tools: tools?.map(OpenAI.toTool),
+ messages: OpenAI.toOpenAIMessage(messages),
+ top_p: this.topP,
+ ...Object.assign({}, this.additionalChatOptions, additionalChatOptions),
+ };
+
+ if (
+ Array.isArray(baseRequestParams.tools) &&
+ baseRequestParams.tools.length === 0
+ ) {
+ // remove empty tools array to avoid OpenAI error
+ delete baseRequestParams.tools;
+ }
+
+ // Streaming
+ if (stream) {
+ return this.streamChat(baseRequestParams);
+ }
+
+ // Non-streaming
+ const response = await this.session.openai.chat.completions.create({
+ ...baseRequestParams,
+ stream: false,
+ });
+
+ const content = response.choices[0]!.message?.content ?? "";
+
+ return {
+ raw: response,
+ message: {
+ content,
+ role: response.choices[0]!.message.role,
+ options: response.choices[0]!.message?.tool_calls
+ ? {
+ toolCall: response.choices[0]!.message.tool_calls.map(
+ (toolCall) => ({
+ id: toolCall.id,
+ name: toolCall.function.name,
+ input: toolCall.function.arguments,
+ }),
+ ),
+ }
+ : {},
+ },
+ };
+ }
+
+ // todo: this wrapper is ugly, refactor it
+ @wrapEventCaller
+ protected async *streamChat(
+ baseRequestParams: OpenAILLM.Chat.ChatCompletionCreateParams,
+ ): AsyncIterable<ChatResponseChunk<ToolCallLLMMessageOptions>> {
+ const stream: AsyncIterable<OpenAILLM.Chat.ChatCompletionChunk> =
+ await this.session.openai.chat.completions.create({
+ ...baseRequestParams,
+ stream: true,
+ });
+
+ // TODO: add callback to streamConverter and use streamConverter here
+ // this will be used to keep track of the current tool call, make sure input are valid json object.
+ let currentToolCall: PartialToolCall | null = null;
+ const toolCallMap = new Map<string, PartialToolCall>();
+ for await (const part of stream) {
+ if (part.choices.length === 0) continue;
+ const choice = part.choices[0]!;
+ // skip parts that don't have any content
+ if (!(choice.delta.content || choice.delta.tool_calls)) continue;
+
+ let shouldEmitToolCall: PartialToolCall | null = null;
+ if (
+ choice.delta.tool_calls?.[0]!.id &&
+ currentToolCall &&
+ choice.delta.tool_calls?.[0].id !== currentToolCall.id
+ ) {
+ shouldEmitToolCall = {
+ ...currentToolCall,
+ input: JSON.parse(currentToolCall.input),
+ };
+ }
+ if (choice.delta.tool_calls?.[0]!.id) {
+ currentToolCall = {
+ name: choice.delta.tool_calls[0].function!.name!,
+ id: choice.delta.tool_calls[0].id,
+ input: choice.delta.tool_calls[0].function!.arguments!,
+ };
+ toolCallMap.set(choice.delta.tool_calls[0].id, currentToolCall);
+ } else {
+ if (choice.delta.tool_calls?.[0]!.function?.arguments) {
+ currentToolCall!.input +=
+ choice.delta.tool_calls[0].function.arguments;
+ }
+ }
+
+ const isDone: boolean = choice.finish_reason !== null;
+
+ if (isDone && currentToolCall) {
+ // for the last one, we need to emit the tool call
+ shouldEmitToolCall = {
+ ...currentToolCall,
+ input: JSON.parse(currentToolCall.input),
+ };
+ }
+
+ yield {
+ raw: part,
+ options: shouldEmitToolCall
+ ? { toolCall: [shouldEmitToolCall] }
+ : currentToolCall
+ ? {
+ toolCall: [currentToolCall],
+ }
+ : {},
+ delta: choice.delta.content ?? "",
+ };
+ }
+ toolCallMap.clear();
+ return;
+ }
+
+ static toTool(tool: BaseTool): ChatCompletionTool {
+ return {
+ type: "function",
+ function: tool.metadata.parameters
+ ? {
+ name: tool.metadata.name,
+ description: tool.metadata.description,
+ parameters: tool.metadata.parameters,
+ }
+ : {
+ name: tool.metadata.name,
+ description: tool.metadata.description,
+ },
+ };
+ }
+}
diff --git a/packages/llm/openai/tsconfig.json b/packages/llm/openai/tsconfig.json
new file mode 100644
index 0000000000000000000000000000000000000000..5a94aa033116766429d5c2d42f337c5b1393ba86
--- /dev/null
+++ b/packages/llm/openai/tsconfig.json
@@ -0,0 +1,18 @@
+{
+ "extends": "../../../tsconfig.json",
+ "compilerOptions": {
+ "target": "ESNext",
+ "module": "ESNext",
+ "moduleResolution": "bundler",
+ "outDir": "./lib"
+ },
+ "include": ["./src"],
+ "references": [
+ {
+ "path": "../../llamaindex/tsconfig.json"
+ },
+ {
+ "path": "../../env/tsconfig.json"
+ }
+ ]
+}
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index c7ed08d7392b6c652f19cfdfa4996fb516f08a10..8186577fe5ea7585961ecaf302f3701a95746da2 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -516,6 +516,9 @@ importers:
'@llamaindex/env':
specifier: workspace:*
version: link:../env
+ '@llamaindex/openai':
+ specifier: workspace:*
+ version: link:../llm/openai
'@mistralai/mistralai':
specifier: ^1.0.4
version: 1.0.4(zod@3.23.8)
@@ -652,9 +655,6 @@ importers:
'@faker-js/faker':
specifier: ^8.4.1
version: 8.4.1
- '@llamaindex/core':
- specifier: workspace:*
- version: link:../../core
'@types/node':
specifier: ^22.5.1
version: 22.5.1
@@ -853,6 +853,25 @@ importers:
specifier: ^2.0.5
version: 2.0.5(@types/node@22.5.1)(terser@5.31.6)
+ packages/llm/openai:
+ dependencies:
+ '@llamaindex/core':
+ specifier: workspace:*
+ version: link:../../core
+ '@llamaindex/env':
+ specifier: workspace:*
+ version: link:../../env
+ openai:
+ specifier: ^4.60.0
+ version: 4.60.0(encoding@0.1.13)(zod@3.23.8)
+ remeda:
+ specifier: ^2.12.0
+ version: 2.12.0
+ devDependencies:
+ bunchee:
+ specifier: 5.3.2
+ version: 5.3.2(typescript@5.5.4)
+
packages/wasm-tools:
dependencies:
'@assemblyscript/loader':
@@ -9472,6 +9491,9 @@ packages:
remark-stringify@11.0.0:
resolution: {integrity: sha512-1OSmLd3awB/t8qdoEOMazZkNsfVTeY4fTsgzcQFdXNq8ToTN4ZGwrMnlda4K6smTFKD+GRV6O48i6Z4iKgPPpw==}
+ remeda@2.12.0:
+ resolution: {integrity: sha512-VAlyhh1os8boCA9/7yN9sXzo0tfCeOwScGXztwBspS0DXQmbIN8xTBfEABvbAW8rMJMPzqxQ1UymHquuESh/pg==}
+
renderkid@3.0.0:
resolution: {integrity: sha512-q/7VIQA8lmM1hF+jn+sFSPWGlMkSAeNYcPLmDQx2zzuiDfaLrOmumR8iaUKlenFgh0XRPIUeSPlH3A+AW3Z5pg==}
@@ -10456,6 +10478,10 @@ packages:
resolution: {integrity: sha512-RAH822pAdBgcNMAfWnCBU3CFZcfZ/i1eZjwFU/dsLKumyuuP3niueg2UAukXYF0E2AAoc82ZSSf9J0WQBinzHA==}
engines: {node: '>=12.20'}
+ type-fest@4.26.1:
+ resolution: {integrity: sha512-yOGpmOAL7CkKe/91I5O3gPICmJNLJ1G4zFYVAsRHg7M64biSnPtRj0WNQt++bRkjYOqjWXrhnUw1utzmVErAdg==}
+ engines: {node: '>=16'}
+
type-is@1.6.18:
resolution: {integrity: sha512-TkRKr9sUTxEH8MdfuCSP7VizJyzRNMjj2J2do2Jr3Kym598JVdEksuzPQCnlFPW4ky9Q+iA+ma9BGm06XQBy8g==}
engines: {node: '>= 0.6'}
@@ -14763,7 +14789,7 @@ snapshots:
'@smithy/is-array-buffer@2.2.0':
dependencies:
- tslib: 2.6.3
+ tslib: 2.7.0
'@smithy/is-array-buffer@3.0.0':
dependencies:
@@ -22166,6 +22192,10 @@ snapshots:
mdast-util-to-markdown: 2.1.0
unified: 11.0.5
+ remeda@2.12.0:
+ dependencies:
+ type-fest: 4.26.1
+
renderkid@3.0.0:
dependencies:
css-select: 4.3.0
@@ -23263,6 +23293,8 @@ snapshots:
type-fest@2.19.0: {}
+ type-fest@4.26.1: {}
+
type-is@1.6.18:
dependencies:
media-typer: 0.3.0
diff --git a/pnpm-workspace.yaml b/pnpm-workspace.yaml
index 79bfed33e0bbc99608bb4c778da032e07c8b5b4d..78e2bf350152e5bf4ec6356c5c525bd84fd26f79 100644
--- a/pnpm-workspace.yaml
+++ b/pnpm-workspace.yaml
@@ -1,6 +1,7 @@
packages:
- "apps/*"
- "packages/*"
+ - "packages/llm/*"
- "packages/core/tests"
- "packages/llamaindex/tests"
- "packages/llamaindex/e2e"
diff --git a/tsconfig.json b/tsconfig.json
index 3a11d68c083c49212e2dd620109e5ac9f4e8b0a6..11288b4933d01f94ba7bc12cf5766135c0079814 100644
--- a/tsconfig.json
+++ b/tsconfig.json
@@ -26,6 +26,9 @@
{
"path": "./packages/community/tsconfig.json"
},
+ {
+ "path": "./packages/llm/openai/tsconfig.json"
+ },
{
"path": "./packages/cloud/tsconfig.json"
},