🤖 feat(Anthropic): Claude 3 & Vision Support (#1984)

* chore: bump anthropic SDK

* chore: update anthropic config settings (fileSupport, default models)

* feat: anthropic multi modal formatting

* refactor: update vision models and use endpoint specific max long side resizing

* feat(anthropic): multimodal messages, retry logic, and messages payload

* chore: add more safety to trimming content due to whitespace error for assistant messages

* feat(anthropic): token accounting and resending multiple images in progress

* chore: bump data-provider

* feat(anthropic): resendImages feature

* chore: optimize Edit/Ask controllers, switch model back to req model

* fix: false positive of invalid model

* refactor(validateVisionModel): use object as arg, pass in additional/available models

* refactor(validateModel): use helper function, `getModelsConfig`

* feat: add modelsConfig to endpointOption so it gets passed to all clients, use for properly validating vision models

* refactor: initialize default vision model and make sure it's available before assigning it

* refactor(useSSE): avoid resetting model if user selected a new model between request and response

* feat: show rate in transaction logging

* fix: return tokenCountMap regardless of payload shape
This commit is contained in:
Danny Avila 2024-03-06 00:04:52 -05:00 committed by GitHub
parent b023c5683d
commit 8263ddda3f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
28 changed files with 599 additions and 115 deletions

View file

@ -4,7 +4,6 @@ const { GoogleVertexAI } = require('langchain/llms/googlevertexai');
const { ChatGoogleGenerativeAI } = require('@langchain/google-genai');
const { ChatGoogleVertexAI } = require('langchain/chat_models/googlevertexai');
const { AIMessage, HumanMessage, SystemMessage } = require('langchain/schema');
const { encodeAndFormat } = require('~/server/services/Files/images');
const { encoding_for_model: encodingForModel, get_encoding: getEncoding } = require('tiktoken');
const {
validateVisionModel,
@ -13,6 +12,7 @@ const {
EModelEndpoint,
AuthKeys,
} = require('librechat-data-provider');
const { encodeAndFormat } = require('~/server/services/Files/images');
const { getModelMaxTokens } = require('~/utils');
const { formatMessage } = require('./prompts');
const BaseClient = require('./BaseClient');
@ -124,18 +124,28 @@ class GoogleClient extends BaseClient {
// stop: modelOptions.stop // no stop method for now
};
if (this.options.attachments) {
this.modelOptions.model = 'gemini-pro-vision';
/* Validation vision request */
this.defaultVisionModel = this.options.visionModel ?? 'gemini-pro-vision';
const availableModels = this.options.modelsConfig?.[EModelEndpoint.google];
this.isVisionModel = validateVisionModel({ model: this.modelOptions.model, availableModels });
if (
this.options.attachments &&
availableModels?.includes(this.defaultVisionModel) &&
!this.isVisionModel
) {
this.modelOptions.model = this.defaultVisionModel;
this.isVisionModel = true;
}
// TODO: as of 12/14/23, only gemini models are "Generative AI" models provided by Google
this.isGenerativeModel = this.modelOptions.model.includes('gemini');
this.isVisionModel = validateVisionModel(this.modelOptions.model);
const { isGenerativeModel } = this;
if (this.isVisionModel && !this.options.attachments) {
this.modelOptions.model = 'gemini-pro';
this.isVisionModel = false;
}
// TODO: as of 12/14/23, only gemini models are "Generative AI" models provided by Google
this.isGenerativeModel = this.modelOptions.model.includes('gemini');
const { isGenerativeModel } = this;
this.isChatModel = !isGenerativeModel && this.modelOptions.model.includes('chat');
const { isChatModel } = this;
this.isTextModel =