Add embedding models configurable, from both transformers.js and TEI (#…

…646)

* Add embedding models configurable, from both Xenova and TEI

* fix lint and format

* Fix bug in sentenceSimilarity

* Batches for TEI using /info route

* Fix web search disapear when finish searching

* Fix lint and format

* Add more options for better embedding model usage

* Fixing CR issues

* Fix websearch disapear in later PR

* Fix lint

* Fix more minor code CR

* Valiadate embeddingModelName field in model config

* Add embeddingModel into shared conversation

* Fix lint and format

* Add default embedding model, and more readme explanation

* Fix minor embedding model readme detailed

* Update settings.json

* Update README.md

Co-authored-by: Mishig <[email protected]>

* Update README.md

Co-authored-by: Mishig <[email protected]>

* Apply suggestions from code review

Co-authored-by: Mishig <[email protected]>

* Resolved more issues

* lint

* Fix more issues

* Fix format

* fix small typo

* lint

* fix default model

* Rn `maxSequenceLength` -> `chunkCharLength`

* format

* add "authorization" example

* format

---------

Co-authored-by: Mishig <[email protected]>
Co-authored-by: Nathan Sarrazin <[email protected]>
Co-authored-by: Mishig Davaadorj <[email protected]>

.env

@@ -46,6 +46,18 @@ CA_PATH=#
 CLIENT_KEY_PASSWORD=#
 REJECT_UNAUTHORIZED=true
 
+TEXT_EMBEDDING_MODELS = `[
+  {
+    "name": "Xenova/gte-small",
+    "displayName": "Xenova/gte-small",
+    "description": "Local embedding model running on the server.",
+    "chunkCharLength": 512,
+    "endpoints": [
+      { "type": "transformersjs" }
+    ]
+  }
+]`
+
 # 'name', 'userMessageToken', 'assistantMessageToken' are required
 MODELS=`[
     {

.env.template

@@ -204,7 +204,6 @@ TASK_MODEL='mistralai/Mistral-7B-Instruct-v0.2'
 #         "stop": ["</s>"]
 # }}`
 
-
 APP_BASE="/chat"
 PUBLIC_ORIGIN=https://huggingface.co
 PUBLIC_SHARE_PREFIX=https://hf.co/chat

README.md

@@ -20,9 +20,10 @@ A chat interface using open source models, eg OpenAssistant or Llama. It is a Sv
 1. [Setup](#setup)
 2. [Launch](#launch)
 3. [Web Search](#web-search)
-4. [Extra parameters](#extra-parameters)
-5. [Deploying to a HF Space](#deploying-to-a-hf-space)
-6. [Building](#building)
+4. [Text Embedding Models](#text-embedding-models)
+5. [Extra parameters](#extra-parameters)
+6. [Deploying to a HF Space](#deploying-to-a-hf-space)
+7. [Building](#building)
 
 ## No Setup Deploy
 
@@ -78,10 +79,50 @@ Chat UI features a powerful Web Search feature. It works by:
 
 1. Generating an appropriate search query from the user prompt.
 2. Performing web search and extracting content from webpages.
-3. Creating embeddings from texts using [transformers.js](https://huggingface.co/docs/transformers.js). Specifically, using [Xenova/gte-small](https://huggingface.co/Xenova/gte-small) model.
+3. Creating embeddings from texts using a text embedding model.
 4. From these embeddings, find the ones that are closest to the user query using a vector similarity search. Specifically, we use `inner product` distance.
 5. Get the corresponding texts to those closest embeddings and perform [Retrieval-Augmented Generation](https://huggingface.co/papers/2005.11401) (i.e. expand user prompt by adding those texts so that an LLM can use this information).
 
+## Text Embedding Models
+
+By default (for backward compatibility), when `TEXT_EMBEDDING_MODELS` environment variable is not defined, [transformers.js](https://huggingface.co/docs/transformers.js) embedding models will be used for embedding tasks, specifically, [Xenova/gte-small](https://huggingface.co/Xenova/gte-small) model.
+
+You can customize the embedding model by setting `TEXT_EMBEDDING_MODELS` in your `.env.local` file. For example:
+
+```env
+TEXT_EMBEDDING_MODELS = `[
+  {
+    "name": "Xenova/gte-small",
+    "displayName": "Xenova/gte-small",
+    "description": "locally running embedding",
+    "chunkCharLength": 512,
+    "endpoints": [
+      {"type": "transformersjs"}
+    ]
+  },
+  {
+    "name": "intfloat/e5-base-v2",
+    "displayName": "intfloat/e5-base-v2",
+    "description": "hosted embedding model",
+    "chunkCharLength": 768,
+    "preQuery": "query: ", # See https://huggingface.co/intfloat/e5-base-v2#faq
+    "prePassage": "passage: ", # See https://huggingface.co/intfloat/e5-base-v2#faq
+    "endpoints": [
+      {
+        "type": "tei",
+        "url": "http://127.0.0.1:8080/",
+        "authorization": "TOKEN_TYPE TOKEN" // optional authorization field. Example: "Basic VVNFUjpQQVNT"
+      }
+    ]
+  }
+]`
+```
+
+The required fields are `name`, `chunkCharLength` and `endpoints`.
+Supported text embedding backends are: [`transformers.js`](https://huggingface.co/docs/transformers.js) and [`TEI`](https://github.com/huggingface/text-embeddings-inference). `transformers.js` models run locally as part of `chat-ui`, whereas `TEI` models run in a different environment & accessed through an API endpoint.
+
+When more than one embedding models are supplied in `.env.local` file, the first will be used by default, and the others will only be used on LLM's which configured `embeddingModel` to the name of the model.
+
 ## Extra parameters
 
 ### OpenID connect
@@ -425,6 +466,45 @@ If you're using a certificate signed by a private CA, you will also need to add
 
 If you're using a self-signed certificate, e.g. for testing or development purposes, you can set the `REJECT_UNAUTHORIZED` parameter to `false` in your `.env.local`. This will disable certificate validation, and allow Chat UI to connect to your custom endpoint.
 
+#### Specific Embedding Model
+
+A model can use any of the embedding models defined in `.env.local`, (currently used when web searching),
+by default it will use the first embedding model, but it can be changed with the field `embeddingModel`:
+
+```env
+TEXT_EMBEDDING_MODELS = `[
+  {
+    "name": "Xenova/gte-small",
+    "chunkCharLength": 512,
+    "endpoints": [
+      {"type": "transformersjs"}
+    ]
+  },
+  {
+    "name": "intfloat/e5-base-v2",
+    "chunkCharLength": 768,
+    "endpoints": [
+      {"type": "tei", "url": "http://127.0.0.1:8080/", "authorization": "Basic VVNFUjpQQVNT"},
+      {"type": "tei", "url": "http://127.0.0.1:8081/"}
+    ]
+  }
+]`
+
+MODELS=`[
+  {
+      "name": "Ollama Mistral",
+      "chatPromptTemplate": "...",
+      "embeddingModel": "intfloat/e5-base-v2"
+      "parameters": {
+        ...
+      },
+      "endpoints": [
+        ...
+      ]
+  }
+]`
+```
+
 ## Deploying to a HF Space
 
 Create a `DOTENV_LOCAL` secret to your HF space with the content of your .env.local, and they will be picked up automatically when you run.

src/lib/components/OpenWebSearchResults.svelte

@@ -30,8 +30,8 @@
 		{:else}
 			<CarbonCheckmark class="my-auto text-gray-500" />
 		{/if}
-		<span class="px-2 font-medium" class:text-red-700={error} class:dark:text-red-500={error}
-			>Web search
+		<span class="px-2 font-medium" class:text-red-700={error} class:dark:text-red-500={error}>
+			Web search
 		</span>
 		<div class="my-auto transition-all" class:rotate-90={detailsOpen}>
 			<CarbonCaretRight />

src/lib/server/embeddingEndpoints/tei/embeddingEndpoints.ts

@@ -0,0 +1,65 @@
+import { z } from "zod";
+import type { EmbeddingEndpoint, Embedding } from "$lib/types/EmbeddingEndpoints";
+import { chunk } from "$lib/utils/chunk";
+
+export const embeddingEndpointTeiParametersSchema = z.object({
+	weight: z.number().int().positive().default(1),
+	model: z.any(),
+	type: z.literal("tei"),
+	url: z.string().url(),
+	authorization: z.string().optional(),
+});
+
+const getModelInfoByUrl = async (url: string, authorization?: string) => {
+	const { origin } = new URL(url);
+
+	const response = await fetch(`${origin}/info`, {
+		headers: {
+			Accept: "application/json",
+			"Content-Type": "application/json",
+			...(authorization ? { Authorization: authorization } : {}),
+		},
+	});
+
+	const json = await response.json();
+	return json;
+};
+
+export async function embeddingEndpointTei(
+	input: z.input<typeof embeddingEndpointTeiParametersSchema>
+): Promise<EmbeddingEndpoint> {
+	const { url, model, authorization } = embeddingEndpointTeiParametersSchema.parse(input);
+
+	const { max_client_batch_size, max_batch_tokens } = await getModelInfoByUrl(url);
+	const maxBatchSize = Math.min(
+		max_client_batch_size,
+		Math.floor(max_batch_tokens / model.chunkCharLength)
+	);
+
+	return async ({ inputs }) => {
+		const { origin } = new URL(url);
+
+		const batchesInputs = chunk(inputs, maxBatchSize);
+
+		const batchesResults = await Promise.all(
+			batchesInputs.map(async (batchInputs) => {
+				const response = await fetch(`${origin}/embed`, {
+					method: "POST",
+					headers: {
+						Accept: "application/json",
+						"Content-Type": "application/json",
+						...(authorization ? { Authorization: authorization } : {}),
+					},
+					body: JSON.stringify({ inputs: batchInputs, normalize: true, truncate: true }),
+				});
+
+				const embeddings: Embedding[] = await response.json();
+				return embeddings;
+			})
+		);
+
+		const flatAllEmbeddings = batchesResults.flat();
+
+		return flatAllEmbeddings;
+	};
+}

src/lib/server/embeddingEndpoints/transformersjs/embeddingEndpoints.ts

@@ -0,0 +1,46 @@
+import { z } from "zod";
+import type { EmbeddingEndpoint } from "$lib/types/EmbeddingEndpoints";
+import type { Tensor, Pipeline } from "@xenova/transformers";
+import { pipeline } from "@xenova/transformers";
+
+export const embeddingEndpointTransformersJSParametersSchema = z.object({
+	weight: z.number().int().positive().default(1),
+	model: z.any(),
+	type: z.literal("transformersjs"),
+});
+
+// Use the Singleton pattern to enable lazy construction of the pipeline.
+class TransformersJSModelsSingleton {
+	static instances: Array<[string, Promise<Pipeline>]> = [];
+
+	static async getInstance(modelName: string): Promise<Pipeline> {
+		const modelPipelineInstance = this.instances.find(([name]) => name === modelName);
+
+		if (modelPipelineInstance) {
+			const [, modelPipeline] = modelPipelineInstance;
+			return modelPipeline;
+		}
+
+		const newModelPipeline = pipeline("feature-extraction", modelName);
+		this.instances.push([modelName, newModelPipeline]);
+
+		return newModelPipeline;
+	}
+}
+
+export async function calculateEmbedding(modelName: string, inputs: string[]) {
+	const extractor = await TransformersJSModelsSingleton.getInstance(modelName);
+	const output: Tensor = await extractor(inputs, { pooling: "mean", normalize: true });
+
+	return output.tolist();
+}
+
+export function embeddingEndpointTransformersJS(
+	input: z.input<typeof embeddingEndpointTransformersJSParametersSchema>
+): EmbeddingEndpoint {
+	const { model } = embeddingEndpointTransformersJSParametersSchema.parse(input);
+
+	return async ({ inputs }) => {
+		return calculateEmbedding(model.name, inputs);
+	};
+}

src/lib/server/embeddingModels.ts

@@ -0,0 +1,99 @@
+import { TEXT_EMBEDDING_MODELS } from "$env/static/private";
+
+import { z } from "zod";
+import { sum } from "$lib/utils/sum";
+import {
+	embeddingEndpoints,
+	embeddingEndpointSchema,
+	type EmbeddingEndpoint,
+} from "$lib/types/EmbeddingEndpoints";
+import { embeddingEndpointTransformersJS } from "$lib/server/embeddingEndpoints/transformersjs/embeddingEndpoints";
+
+const modelConfig = z.object({
+	/** Used as an identifier in DB */
+	id: z.string().optional(),
+	/** Used to link to the model page, and for inference */
+	name: z.string().min(1),
+	displayName: z.string().min(1).optional(),
+	description: z.string().min(1).optional(),
+	websiteUrl: z.string().url().optional(),
+	modelUrl: z.string().url().optional(),
+	endpoints: z.array(embeddingEndpointSchema).nonempty(),
+	chunkCharLength: z.number().positive(),
+	preQuery: z.string().default(""),
+	prePassage: z.string().default(""),
+});
+
+// Default embedding model for backward compatibility
+const rawEmbeddingModelJSON =
+	TEXT_EMBEDDING_MODELS ||
+	`[
+	{
+	  "name": "Xenova/gte-small",
+	  "chunkCharLength": 512,
+	  "endpoints": [
+		{ "type": "transformersjs" }
+	  ]
+	}
+]`;
+
+const embeddingModelsRaw = z.array(modelConfig).parse(JSON.parse(rawEmbeddingModelJSON));
+
+const processEmbeddingModel = async (m: z.infer<typeof modelConfig>) => ({
+	...m,
+	id: m.id || m.name,
+});
+
+const addEndpoint = (m: Awaited<ReturnType<typeof processEmbeddingModel>>) => ({
+	...m,
+	getEndpoint: async (): Promise<EmbeddingEndpoint> => {
+		if (!m.endpoints) {
+			return embeddingEndpointTransformersJS({
+				type: "transformersjs",
+				weight: 1,
+				model: m,
+			});
+		}
+
+		const totalWeight = sum(m.endpoints.map((e) => e.weight));
+
+		let random = Math.random() * totalWeight;
+
+		for (const endpoint of m.endpoints) {
+			if (random < endpoint.weight) {
+				const args = { ...endpoint, model: m };
+
+				switch (args.type) {
+					case "tei":
+						return embeddingEndpoints.tei(args);
+					case "transformersjs":
+						return embeddingEndpoints.transformersjs(args);
+				}
+			}
+
+			random -= endpoint.weight;
+		}
+
+		throw new Error(`Failed to select embedding endpoint`);
+	},
+});
+
+export const embeddingModels = await Promise.all(
+	embeddingModelsRaw.map((e) => processEmbeddingModel(e).then(addEndpoint))
+);
+
+export const defaultEmbeddingModel = embeddingModels[0];
+
+const validateEmbeddingModel = (_models: EmbeddingBackendModel[], key: "id" | "name") => {
+	return z.enum([_models[0][key], ..._models.slice(1).map((m) => m[key])]);
+};
+
+export const validateEmbeddingModelById = (_models: EmbeddingBackendModel[]) => {
+	return validateEmbeddingModel(_models, "id");
+};
+
+export const validateEmbeddingModelByName = (_models: EmbeddingBackendModel[]) => {
+	return validateEmbeddingModel(_models, "name");
+};
+
+export type EmbeddingBackendModel = typeof defaultEmbeddingModel;

src/lib/server/models.ts

@@ -12,6 +12,7 @@ import { z } from "zod";
 import endpoints, { endpointSchema, type Endpoint } from "./endpoints/endpoints";
 import endpointTgi from "./endpoints/tgi/endpointTgi";
 import { sum } from "$lib/utils/sum";
+import { embeddingModels, validateEmbeddingModelByName } from "./embeddingModels";
 
 import JSON5 from "json5";
 
@@ -68,6 +69,7 @@ const modelConfig = z.object({
 		.optional(),
 	multimodal: z.boolean().default(false),
 	unlisted: z.boolean().default(false),
+	embeddingModel: validateEmbeddingModelByName(embeddingModels).optional(),
 });
 
 const modelsRaw = z.array(modelConfig).parse(JSON5.parse(MODELS));