Repository files navigation

Find the docs athf.co/docs/chat-ui.

A chat interface using open source models, eg OpenAssistant or Llama. It is a SvelteKit app and it powers theHuggingChat app on hf.co/chat.

0. Quickstart
1. No Setup Deploy
2. Setup
3. Launch
4. Web Search
5. Text Embedding Models
6. Extra parameters
7. Common issues
8. Deploying to a HF Space
9. Building

You can deploy a chat-ui instance in a single command using the docker image. Get your huggingface token fromhere.

docker run -p 3000 -e HF_TOKEN=hf_*** -v db:/data ghcr.io/huggingface/chat-ui-db:latest

Take a look at the.env file and the readme to see all the environment variables that you can set. We have endpoint support for all OpenAI API compatible local services as well as many other providers like Anthropic, Cloudflare, Google Vertex AI, etc.

You can quickly start a locally running chat-ui & LLM text-generation server thanks to chat-ui'sllama.cpp server support.

Step 1 (Start llama.cpp server):

Install llama.cpp w/ brew (for Mac):

# install llama.cppbrew install llama.cpp

orbuild directly from the source for your target device:

git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp && make

Next, start the server with theLLM of your choice:

# start llama.cpp server (using hf.co/microsoft/Phi-3-mini-4k-instruct-gguf as an example)llama-server --hf-repo microsoft/Phi-3-mini-4k-instruct-gguf --hf-file Phi-3-mini-4k-instruct-q4.gguf -c 4096

A local LLaMA.cpp HTTP Server will start onhttp://localhost:8080. Read morehere.

Step 3 (make sure you have MongoDb running locally):

docker run -d -p 27017:27017 --name mongo-chatui mongo:latest

Read morehere.

Step 4 (clone chat-ui):

git clone https://github.com/huggingface/chat-uicd chat-ui

Step 5 (tell chat-ui to use local llama.cpp server):

Add the following to your.env.local:

MODELS=`[  {"name":"microsoft/Phi-3-mini-4k-instruct","endpoints": [{"type" :"llamacpp","baseURL":"http://localhost:8080"    }],  },]`

Read morehere.

Step 6 (start chat-ui):

npm installnpm run dev -- --open

Read morehere.

If you don't want to configure, setup, and launch your own Chat UI yourself, you can use this option as a fast deploy alternative.

You can deploy your own customized Chat UI instance with any supportedLLM of your choice onHugging Face Spaces. To do so, use the chat-ui templateavailable here.

SetHF_TOKEN inSpace secrets to deploy a model with gated access or a model in a private repository. It's also compatible withInference for PROs curated list of powerful models with higher rate limits. Make sure to create your personal token first in yourUser Access Tokens settings.

Read the full tutorialhere.

The default config for Chat UI is stored in the.env file. You will need to override some values to get Chat UI to run locally. This is done in.env.local.

Start by creating a.env.local file in the root of the repository. The bare minimum config you need to get Chat UI to run locally is the following:

MONGODB_URL=<the URL to your MongoDB instance>HF_TOKEN=<your access token>

The chat history is stored in a MongoDB instance, and having a DB instance available is needed for Chat UI to work.

You can use a local MongoDB instance. The easiest way is to spin one up using docker:

docker run -d -p 27017:27017 --name mongo-chatui mongo:latest

In which case the url of your DB will beMONGODB_URL=mongodb://localhost:27017.

Alternatively, you can use afree MongoDB Atlas instance for this, Chat UI should fit comfortably within their free tier. After which you can set theMONGODB_URL variable in.env.local to match your instance.

If you use a remote inference endpoint, you will need a Hugging Face access token to run Chat UI locally. You can get one fromyour Hugging Face profile.

After you're done with the.env.local file you can run Chat UI locally with:

npm installnpm run dev

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 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 useinner product distance.
5. Get the corresponding texts to those closest embeddings and performRetrieval-Augmented Generation (i.e. expand user prompt by adding those texts so that an LLM can use this information).

By default (for backward compatibility), whenTEXT_EMBEDDING_MODELS environment variable is not defined,transformers.js embedding models will be used for embedding tasks, specifically,Xenova/gte-small model.

You can customize the embedding model by settingTEXT_EMBEDDING_MODELS in your.env.local file. For example:

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 arename,chunkCharLength andendpoints.Supported text embedding backends are:transformers.js,TEI andOpenAI.transformers.js models run locally as part ofchat-ui, whereasTEI models run in a different environment & accessed through an API endpoint.openai models are accessed through theOpenAI API.

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 configuredembeddingModel to the name of the model.

The login feature is disabled by default and users are attributed a unique ID based on their browser. But if you want to use OpenID to authenticate your users, you can add the following to your.env.local file:

OPENID_CONFIG=`{PROVIDER_URL: "<your OIDC issuer>",CLIENT_ID: "<your OIDC client ID>",CLIENT_SECRET: "<your OIDC client secret>",SCOPES: "openid profile",TOLERANCE:// optionalRESOURCE:// optional}`

These variables will enable the openID sign-in modal for users.

You can set the env variableTRUSTED_EMAIL_HEADER to point to the header that contains the user's email address. This will allow you to authenticate users from the header. This setup is usually combined with a proxy that will be in front of chat-ui and will handle the auth and set the header.

Here is a list of header names for common auth providers:

• Tailscale Serve:Tailscale-User-Login
• Cloudflare Access:Cf-Access-Authenticated-User-Email
• oauth2-proxy:X-Forwarded-Email

You can use a few environment variables to customize the look and feel of chat-ui. These are by default:

PUBLIC_APP_NAME=ChatUIPUBLIC_APP_ASSETS=chatuiPUBLIC_APP_COLOR=bluePUBLIC_APP_DESCRIPTION="Making the community's best AI chat models available to everyone."PUBLIC_APP_DATA_SHARING=PUBLIC_APP_DISCLAIMER=

• PUBLIC_APP_NAME The name used as a title throughout the app.
• PUBLIC_APP_ASSETS Is used to find logos & favicons instatic/$PUBLIC_APP_ASSETS, current options arechatui andhuggingchat.
• PUBLIC_APP_COLOR Can be any of thetailwind colors.
• PUBLIC_APP_DATA_SHARING Can be set to 1 to add a toggle in the user settings that lets your users opt-in to data sharing with models creator.
• PUBLIC_APP_DISCLAIMER If set to 1, we show a disclaimer about generated outputs on login.

You can enable the web search through an API by addingYDC_API_KEY (docs.you.com) orSERPER_API_KEY (serper.dev) orSERPAPI_KEY (serpapi.com) orSERPSTACK_API_KEY (serpstack.com) orSEARCHAPI_KEY (searchapi.io) to your.env.local.

You can also simply enable the local google websearch by settingUSE_LOCAL_WEBSEARCH=true in your.env.local or specify a SearXNG instance by adding the query URL toSEARXNG_QUERY_URL.

You can enable javascript when parsing webpages to improve compatibility withWEBSEARCH_JAVASCRIPT=true at the cost of increased CPU usage. You'll want at least 4 cores when enabling.

You can customize the parameters passed to the model or even use a new model by updating theMODELS variable in your.env.local. The default one can be found in.env and looks like this :

MODELS=`[  {    "name": "mistralai/Mistral-7B-Instruct-v0.2",    "displayName": "mistralai/Mistral-7B-Instruct-v0.2",    "description": "Mistral 7B is a new Apache 2.0 model, released by Mistral AI that outperforms Llama2 13B in benchmarks.",    "websiteUrl": "https://mistral.ai/news/announcing-mistral-7b/",    "preprompt": "",    "chatPromptTemplate" : "<s>{{#each messages}}{{#ifUser}}[INST] {{#if @first}}{{#if @root.preprompt}}{{@root.preprompt}}\n{{/if}}{{/if}}{{content}} [/INST]{{/ifUser}}{{#ifAssistant}}{{content}}</s>{{/ifAssistant}}{{/each}}",    "parameters": {      "temperature": 0.3,      "top_p": 0.95,      "repetition_penalty": 1.2,      "top_k": 50,      "truncate": 3072,      "max_new_tokens": 1024,      "stop": ["</s>"]    },    "promptExamples": [      {        "title": "Write an email from bullet list",        "prompt": "As a restaurant owner, write a professional email to the supplier to get these products every week: \n\n- Wine (x10)\n- Eggs (x24)\n- Bread (x12)"      }, {        "title": "Code a snake game",        "prompt": "Code a basic snake game in python, give explanations for each step."      }, {        "title": "Assist in a task",        "prompt": "How do I make a delicious lemon cheesecake?"      }    ]  }]`

You can change things like the parameters, or customize the preprompt to better suit your needs. You can also add more models by adding more objects to the array, with different preprompts for example.

When querying the model for a chat response, thechatPromptTemplate template is used.messages is an array of chat messages, it has the format[{ content: string }, ...]. To identify if a message is a user message or an assistant message theifUser andifAssistant block helpers can be used.

The following is the defaultchatPromptTemplate, although newlines and indentiation have been added for readability. You can find the prompts used in production for HuggingChathere.

{{preprompt}}{{#each messages}}  {{#ifUser}}{{@root.userMessageToken}}{{content}}{{@root.userMessageEndToken}}{{/ifUser}}  {{#ifAssistant}}{{@root.assistantMessageToken}}{{content}}{{@root.assistantMessageEndToken}}{{/ifAssistant}}{{/each}}{{assistantMessageToken}}

[!INFO]We also support Jinja2 templates for thechatPromptTemplate in addition to Handlebars templates. On startup we first try to compile with Jinja and if that fails we fall back to interpretingchatPromptTemplate as handlebars.

We currently supportIDEFICS (hosted on TGI), OpenAI and Claude 3 as multimodal models. You can enable it by settingmultimodal: true in yourMODELS configuration. For IDEFICS, you must have aPRO HF Api token. For OpenAI, see theOpenAI section. For Anthropic, see theAnthropic section.

    {      "name": "HuggingFaceM4/idefics-80b-instruct",      "multimodal" : true,      "description": "IDEFICS is the new multimodal model by Hugging Face.",      "preprompt": "",      "chatPromptTemplate" : "{{#each messages}}{{#ifUser}}User: {{content}}{{/ifUser}}<end_of_utterance>\nAssistant: {{#ifAssistant}}{{content}}\n{{/ifAssistant}}{{/each}}",      "parameters": {        "temperature": 0.1,        "top_p": 0.95,        "repetition_penalty": 1.2,        "top_k": 12,        "truncate": 1000,        "max_new_tokens": 1024,        "stop": ["<end_of_utterance>", "User:", "\nUser:"]      }    }

If you want to, instead of hitting models on the Hugging Face Inference API, you can run your own models locally.

A good option is to hit atext-generation-inference endpoint. This is what is done in the officialChat UI Spaces Docker template for instance: both this app and a text-generation-inference server run inside the same container.

To do this, you can add your own endpoints to theMODELS variable in.env.local, by adding an"endpoints" key for each model inMODELS.

{// rest of the model config here"endpoints": [{  "type" : "tgi",  "url": "https://HOST:PORT",  }]}

Ifendpoints are left unspecified, ChatUI will look for the model on the hosted Hugging Face inference API using the model name.

Chat UI can be used with any API server that supports OpenAI API compatibility, for exampletext-generation-webui,LocalAI,FastChat,llama-cpp-python, andialacol andvllm.

The following example config makes Chat UI works withtext-generation-webui, theendpoint.baseUrl is the url of the OpenAI API compatible server, this overrides the baseUrl to be used by OpenAI instance. Theendpoint.completion determine which endpoint to be used, default ischat_completions which usesv1/chat/completions, change toendpoint.completion tocompletions to use thev1/completions endpoint.

Parameters not supported by OpenAI (e.g., top_k, repetition_penalty, etc.) must be set in the extraBody of endpoints. Be aware that setting them in parameters will cause them to be omitted.

MODELS=`[  {    "name": "text-generation-webui",    "id": "text-generation-webui",    "parameters": {      "temperature": 0.9,      "top_p": 0.95,      "max_new_tokens": 1024,      "stop": []    },    "endpoints": [{      "type" : "openai",      "baseURL": "http://localhost:8000/v1",      "extraBody": {        "repetition_penalty": 1.2,        "top_k": 50,        "truncate": 1000      }    }]  }]`

Theopenai type includes official OpenAI models. You can add, for example, GPT4/GPT3.5 as a "openai" model:

OPENAI_API_KEY=#your openai api key hereMODELS=`[{      "name": "gpt-4",      "displayName": "GPT 4",      "endpoints" : [{        "type": "openai"      }]},      {      "name": "gpt-3.5-turbo",      "displayName": "GPT 3.5 Turbo",      "endpoints" : [{        "type": "openai"      }]}]`

You may also consume any model provider that provides compatible OpenAI API endpoint. For example, you may self-hostPortkey gateway and experiment with Claude or GPTs offered by Azure OpenAI. Example for Claude from Anthropic:

MODELS=`[{  "name": "claude-2.1",  "displayName": "Claude 2.1",  "description": "Anthropic has been founded by former OpenAI researchers...",  "parameters": {      "temperature": 0.5,      "max_new_tokens": 4096,  },  "endpoints": [      {          "type": "openai",          "baseURL": "https://gateway.example.com/v1",          "defaultHeaders": {              "x-portkey-config": '{"provider":"anthropic","api_key":"sk-ant-abc...xyz"}'          }      }  ]}]`

Example for GPT 4 deployed on Azure OpenAI:

MODELS=`[{  "id": "gpt-4-1106-preview",  "name": "gpt-4-1106-preview",  "displayName": "gpt-4-1106-preview",  "parameters": {      "temperature": 0.5,      "max_new_tokens": 4096,  },  "endpoints": [      {          "type": "openai",          "baseURL": "https://{resource-name}.openai.azure.com/openai/deployments/{deployment-id}",          "defaultHeaders": {              "api-key": "{api-key}"          },          "defaultQuery": {              "api-version": "2023-05-15"          }      }  ]}]`

Or try Mistral fromDeepinfra:

Note, apiKey can either be set custom per endpoint, or globally usingOPENAI_API_KEY variable.

MODELS=`[{  "name": "mistral-7b",  "displayName": "Mistral 7B",  "description": "A 7B dense Transformer, fast-deployed and easily customisable. Small, yet powerful for a variety of use cases. Supports English and code, and a 8k context window.",  "parameters": {      "temperature": 0.5,      "max_new_tokens": 4096,  },  "endpoints": [      {          "type": "openai",          "baseURL": "https://api.deepinfra.com/v1/openai",          "apiKey": "abc...xyz"      }  ]}]`

Non-streaming endpoints

For endpoints that don´t support streaming like o1 on Azure, you can passstreamingSupported: false in your endpoint config:

MODELS=`[{  "id": "o1-preview",  "name": "o1-preview",  "displayName": "o1-preview",  "systemRoleSupported": false,  "endpoints": [    {      "type": "openai",      "baseURL": "https://my-deployment.openai.azure.com/openai/deployments/o1-preview",      "defaultHeaders": {        "api-key": "$SECRET"      },      "streamingSupported": false,    }  ]}]`

chat-ui also supports the llama.cpp API server directly without the need for an adapter. You can do this using thellamacpp endpoint type.

If you want to run Chat UI with llama.cpp, you can do the following, usingmicrosoft/Phi-3-mini-4k-instruct-gguf as an example model:

# install llama.cppbrew install llama.cpp# start llama.cpp serverllama-server --hf-repo microsoft/Phi-3-mini-4k-instruct-gguf --hf-file Phi-3-mini-4k-instruct-q4.gguf -c 4096

MODELS=`[  {      "name": "Local Zephyr",      "chatPromptTemplate": "<|system|>\n{{preprompt}}</s>\n{{#each messages}}{{#ifUser}}<|user|>\n{{content}}</s>\n<|assistant|>\n{{/ifUser}}{{#ifAssistant}}{{content}}</s>\n{{/ifAssistant}}{{/each}}",      "parameters": {        "temperature": 0.1,        "top_p": 0.95,        "repetition_penalty": 1.2,        "top_k": 50,        "truncate": 1000,        "max_new_tokens": 2048,        "stop": ["</s>"]      },      "endpoints": [        {         "url": "http://127.0.0.1:8080",         "type": "llamacpp"        }      ]  }]`

Start chat-ui withnpm run dev and you should be able to chat with Zephyr locally.

We also support the Ollama inference server. Spin up a model with

ollama run mistral

Then specify the endpoints like so:

MODELS=`[  {      "name": "Ollama Mistral",      "chatPromptTemplate": "<s>{{#each messages}}{{#ifUser}}[INST] {{#if @first}}{{#if @root.preprompt}}{{@root.preprompt}}\n{{/if}}{{/if}} {{content}} [/INST]{{/ifUser}}{{#ifAssistant}}{{content}}</s> {{/ifAssistant}}{{/each}}",      "parameters": {        "temperature": 0.1,        "top_p": 0.95,        "repetition_penalty": 1.2,        "top_k": 50,        "truncate": 3072,        "max_new_tokens": 1024,        "stop": ["</s>"]      },      "endpoints": [        {         "type": "ollama",         "url" : "http://127.0.0.1:11434",         "ollamaName" : "mistral"        }      ]  }]`

We also support Anthropic models (including multimodal ones viamultmodal: true) through the official SDK. You may provide your API key via theANTHROPIC_API_KEY env variable, or alternatively, through theendpoints.apiKey as per the following example.

MODELS=`[  {      "name": "claude-3-haiku-20240307",      "displayName": "Claude 3 Haiku",      "description": "Fastest and most compact model for near-instant responsiveness",      "multimodal": true,      "parameters": {        "max_new_tokens": 4096,      },      "endpoints": [        {          "type": "anthropic",          // optionals          "apiKey": "sk-ant-...",          "baseURL": "https://api.anthropic.com",          "defaultHeaders": {},          "defaultQuery": {}        }      ]  },  {      "name": "claude-3-sonnet-20240229",      "displayName": "Claude 3 Sonnet",      "description": "Ideal balance of intelligence and speed",      "multimodal": true,      "parameters": {        "max_new_tokens": 4096,      },      "endpoints": [        {          "type": "anthropic",          // optionals          "apiKey": "sk-ant-...",          "baseURL": "https://api.anthropic.com",          "defaultHeaders": {},          "defaultQuery": {}        }      ]  },  {      "name": "claude-3-opus-20240229",      "displayName": "Claude 3 Opus",      "description": "Most powerful model for highly complex tasks",      "multimodal": true,      "parameters": {         "max_new_tokens": 4096      },      "endpoints": [        {          "type": "anthropic",          // optionals          "apiKey": "sk-ant-...",          "baseURL": "https://api.anthropic.com",          "defaultHeaders": {},          "defaultQuery": {}        }      ]  }]`

We also support using Anthropic models running on Vertex AI. Authentication is done using Google Application Default Credentials. Project ID can be provided through theendpoints.projectId as per the following example:

MODELS=`[  {      "name": "claude-3-sonnet@20240229",      "displayName": "Claude 3 Sonnet",      "description": "Ideal balance of intelligence and speed",      "multimodal": true,      "parameters": {        "max_new_tokens": 4096,      },      "endpoints": [        {          "type": "anthropic-vertex",          "region": "us-central1",          "projectId": "gcp-project-id",          // optionals          "defaultHeaders": {},          "defaultQuery": {}        }      ]  },  {      "name": "claude-3-haiku@20240307",      "displayName": "Claude 3 Haiku",      "description": "Fastest, most compact model for near-instant responsiveness",      "multimodal": true,      "parameters": {         "max_new_tokens": 4096      },      "endpoints": [        {          "type": "anthropic-vertex",          "region": "us-central1",          "projectId": "gcp-project-id",          // optionals          "defaultHeaders": {},          "defaultQuery": {}        }      ]  }]`

You can also specify your Amazon SageMaker instance as an endpoint for chat-ui. The config goes like this:

"endpoints": [    {      "type" : "aws",      "service" : "sagemaker"      "url": "",      "accessKey": "",      "secretKey" : "",      "sessionToken": "",      "region": "",      "weight": 1    }]

You can also set"service" : "lambda" to use a lambda instance.

You can get theaccessKey andsecretKey from your AWS user, under programmatic access.

You can also use Cloudflare Workers AI to run your own models with serverless inference.

You will need to have a Cloudflare account, then get youraccount ID as well as yourAPI token for Workers AI.

You can either specify them directly in your.env.local using theCLOUDFLARE_ACCOUNT_ID andCLOUDFLARE_API_TOKEN variables, or you can set them directly in the endpoint config.

You can find the list of models available on Cloudflarehere.

  {  "name" : "nousresearch/hermes-2-pro-mistral-7b",  "tokenizer": "nousresearch/hermes-2-pro-mistral-7b",  "parameters": {    "stop": ["<|im_end|>"]  },  "endpoints" : [    {      "type" : "cloudflare"      <!-- optionally specify these      "accountId": "your-account-id",      "authToken": "your-api-token"      -->    }  ]}

You can also use Cohere to run their models directly from chat-ui. You will need to have a Cohere account, then get yourAPI token. You can either specify it directly in your.env.local using theCOHERE_API_TOKEN variable, or you can set it in the endpoint config.

Here is an example of a Cohere model config. You can set which model you want to use by setting theid field to the model name.

  {    "name" : "CohereForAI/c4ai-command-r-v01",    "id": "command-r",    "description": "C4AI Command-R is a research release of a 35 billion parameter highly performant generative model",    "endpoints": [      {        "type": "cohere",        <!-- optionally specify these, or use COHERE_API_TOKEN        "apiKey": "your-api-token"        -->      }    ]  }

Chat UI can connect to the google Vertex API endpoints (List of supported models).

To enable:

1. Select orcreate a Google Cloud project.
2. Enable billing for your project.
3. Enable the Vertex AI API.
4. Set up authentication with a service accountso you can access the API from your local workstation.

The service account credentials file can be imported as an environmental variable:

GOOGLE_APPLICATION_CREDENTIALS=clientid.json

Make sure your docker container has access to the file and the variable is correctly set.Afterwards Google Vertex endpoints can be configured as following:

MODELS=`[//...    {       "name": "gemini-1.5-pro",       "displayName": "Vertex Gemini Pro 1.5",       "multimodal": true,       "endpoints" : [{          "type": "vertex",          "project": "abc-xyz",          "location": "europe-west3",          "extraBody": {          "model_version": "gemini-1.5-pro-preview-0409",          },          // Optional          "safetyThreshold": "BLOCK_MEDIUM_AND_ABOVE",          "apiEndpoint": "", // alternative api endpoint url,          "tools": [{            "googleSearchRetrieval": {              "disableAttribution": true            }          }],          "multimodal": {            "image": {              "supportedMimeTypes": ["image/png", "image/jpeg", "image/webp"],              "preferredMimeType": "image/png",              "maxSizeInMB": 5,              "maxWidth": 2000,              "maxHeight": 1000,            }          }       }]     },]`

LangChain applications that are deployed using LangServe can be called with the following config:

MODELS=`[//...    {       "name": "summarization-chain", //model-name       "endpoints" : [{         "type": "langserve",         "url" : "http://127.0.0.1:8100",       }]     },]`

Custom endpoints may require authorization, depending on how you configure them. Authentication will usually be set either withBasic orBearer.

ForBasic we will need to generate a base64 encoding of the username and password.

echo -n "USER:PASS" | base64

VVNFUjpQQVNT

ForBearer you can use a token, which can be grabbed fromhere.

You can then add the generated information and theauthorization parameter to your.env.local.

"endpoints": [  {    "url": "https://HOST:PORT",    "authorization": "Basic VVNFUjpQQVNT",  }]

Please note that ifHF_TOKEN is also set or not empty, it will take precedence.

If the model being hosted will be available on multiple servers/instances add theweight parameter to your.env.local. Theweight will be used to determine the probability of requesting a particular endpoint.

"endpoints": [  {    "url": "https://HOST:PORT",    "weight": 1  },  {    "url": "https://HOST:PORT",    "weight": 2  }  ...]

Custom endpoints may require client certificate authentication, depending on how you configure them. To enable mTLS between Chat UI and your custom endpoint, you will need to set theUSE_CLIENT_CERTIFICATE totrue, and add theCERT_PATH andKEY_PATH parameters to your.env.local. These parameters should point to the location of the certificate and key files on your local machine. The certificate and key files should be in PEM format. The key file can be encrypted with a passphrase, in which case you will also need to add theCLIENT_KEY_PASSWORD parameter to your.env.local.

If you're using a certificate signed by a private CA, you will also need to add theCA_PATH parameter to your.env.local. This parameter should point to the location of the CA certificate file on your local machine.

If you're using a self-signed certificate, e.g. for testing or development purposes, you can set theREJECT_UNAUTHORIZED parameter tofalse in your.env.local. This will disable certificate validation, and allow Chat UI to connect to your custom endpoint.

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 fieldembeddingModel:

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": [        ...      ]  }]`

ChatUI supports specialized reasoning/Chain-of-Thought (CoT) models through thereasoning configuration field. When properly configured, this displays a UI widget that allows users to view or collapse the model’s reasoning steps. We support three types of reasoning parsing:

For models like DeepSeek R1, token-based delimitations can be used to identify reasoning steps. This is done by specifying thebeginToken andendToken fields in thereasoning configuration.

Example configuration for DeepSeek R1 (token-based):

{"name":"deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",// ..."reasoning": {"type":"tokens","beginToken":"<think>","endToken":"</think>"}}

For models like QwQ, which return a chain of thought but do not explicitly provide a final answer, thesummarize type can be used. This automatically summarizes the reasoning steps using theTASK_MODEL (or the first model in the configuration ifTASK_MODEL is not specified) and displays the summary as the final answer.

Example configuration for QwQ (summarize-based):

{"name":"Qwen/QwQ-32B-Preview",// ..."reasoning": {"type":"summarize"}}

In some cases, the final answer can be extracted from the model output using a regular expression. This is achieved by specifying theregex field in thereasoning configuration. For example, if your model wraps the final answer in a\boxed{} tag, you can use the following configuration:

{"name":"model/yourmodel",// ..."reasoning": {"type":"regex","regex":"\\\\boxed\\{(.+?)\\}"}}

Most likely you are running chat-ui over HTTP. The recommended option is to setup something like NGINX to handle HTTPS and proxy the requests to chat-ui. If you really need to run over HTTP you can addCOOKIE_SECURE=false andCOOKIE_SAMESITE=lax to your.env.local.

Make sure to set yourPUBLIC_ORIGIN in your.env.local to the correct URL as well.

Create aDOTENV_LOCAL secret to your HF space with the content of your .env.local, and they will be picked up automatically when you run.

To create a production version of your app:

npm run build

You can preview the production build withnpm run preview.

To deploy your app, you may need to install anadapter for your target environment.

The config file for HuggingChat is stored in thechart/env/prod.yaml file. It is the source of truth for the environment variables used for our CI/CD pipeline. For HuggingChat, as we need to customize the app color, as well as the base path, we build a custom docker image. You can find the workflow here.

If you want to run an exact copy of HuggingChat locally, you will need to do the following first:

1. Create anOAuth App on the hub withopenid profile email permissions. Make sure to set the callback URL to something likehttp://localhost:5173/chat/login/callback which matches the right path for your local instance.
2. Create aHF Token with your Hugging Face account. You will need a Pro account to be able to access some of the larger models available through HuggingChat.
3. Create a free account withserper.dev (you will get 2500 free search queries)
4. Run an instance of mongoDB, however you want. (Local or remote)

You can then create a new.env.SECRET_CONFIG file with the following content

MONGODB_URL=<link to your mongo DB from step 4>HF_TOKEN=<your HF token from step 2>OPENID_CONFIG=`{PROVIDER_URL: "https://huggingface.co",CLIENT_ID: "<your client ID from step 1>",CLIENT_SECRET: "<your client secret from step 1>",}`SERPER_API_KEY=<your serper API key from step 3>MESSAGES_BEFORE_LOGIN=<can be any numerical value, or set to 0 to require login>

You can then runnpm run updateLocalEnv in the root of chat-ui. This will create a.env.local file which combines thechart/env/prod.yaml and the.env.SECRET_CONFIG file. You can then runnpm run dev to start your local instance of HuggingChat.

You can populate the database using faker data using thepopulate script:

npm run populate<flags here>

At least one flag must be specified, the following flags are available:

• reset - resets the database
• all - populates all tables
• users - populates the users table
• settings - populates the settings table for existing users
• assistants - populates the assistants table for existing users
• conversations - populates the conversations table for existing users

For example, you could use it like so:

npm run populate reset

to clear out the database. Then login in the app to create your user and run the following command:

npm run populate users settings assistants conversations

to populate the database with fake data, including fake conversations and assistants for your user.

You can build the docker images locally using the following commands:

docker build -t chat-ui-db:latest --build-arg INCLUDE_DB=true.docker build -t chat-ui:latest --build-arg INCLUDE_DB=false.docker build -t huggingchat:latest --build-arg INCLUDE_DB=false --build-arg APP_BASE=/chat --build-arg PUBLIC_APP_COLOR=yellow.

If you want to run the images with your local .env.local you have two options

DOTENV_LOCAL=$(<.env.local)  docker run --rm --network=host -e DOTENV_LOCAL -p 3000:3000 chat-ui

docker run --rm --network=host --mount type=bind,source="$(pwd)/.env.local",target=/app/.env.local -p 3000:3000 chat-ui