|
| 1 | +# Conversation building block |
| 2 | + |
| 3 | +* Author(s): Loong Dai (@daixiang0) |
| 4 | +* Updated: 2024-06-18 |
| 5 | + |
| 6 | +## Overview |
| 7 | + |
| 8 | +This is a proposal for a new building block for Dapr to allow developers to leverage LLM services in a consistent way. Goal is to expose an API that allows developers to ask Dapr to do request. |
| 9 | + |
| 10 | +## Background |
| 11 | + |
| 12 | +Now there are many large language model servers or toolkits, which provides own APIs, like [OpenAI](https://openai.com/), [Hugging Face](https://huggingface.co/), [Kserve](https://kserve.github.io/website/latest/), [OpenVINO](https://docs.openvino.ai/) and so on. |
| 13 | + |
| 14 | +For developers, it is hard to migrate from one to the other due to hardcode and API differences. |
| 15 | + |
| 16 | +For startups and communities, they need to implement the popular APIs as soon as possible, or users maybe give up tring or adopting because of the cost of migration. |
| 17 | + |
| 18 | +This is an area where Dapr can help. We can offer an abstraction layer on those APIs. |
| 19 | + |
| 20 | +## Component YAML |
| 21 | + |
| 22 | +A component can have it's own set of attributes, like in Dapr. For example: |
| 23 | + |
| 24 | +```yaml |
| 25 | +apiVersion: dapr.io/v1alpha1 |
| 26 | +kind: Component |
| 27 | +metadata: |
| 28 | + name: chatgpt4o |
| 29 | +spec: |
| 30 | + type: conversation.chatgpt |
| 31 | + version: v1 |
| 32 | + metadata: |
| 33 | + - name: key |
| 34 | + value: "bfnskdlgdhklhk53adfgsfnsgmtyqdghbid34891" |
| 35 | + - name: model |
| 36 | + value: "gpt-4o" |
| 37 | + - name: endpoints |
| 38 | + value: "us.api.openai.com,eu.api.openai.com" |
| 39 | + - name: loadBalancingPolicy |
| 40 | + value: "ROUNDROBIN" |
| 41 | +``` |
| 42 | +
|
| 43 | +## gRPC APIs |
| 44 | +
|
| 45 | +In the Dapr gRPC APIs, we are extending the `runtime.v1.Dapr` service to add new methods: |
| 46 | + |
| 47 | +> Note: APIs will have "Alpha1" added while in preview |
| 48 | + |
| 49 | +> Note: The API token is stored in the component |
| 50 | + |
| 51 | +```proto |
| 52 | +// (Existing Dapr service) |
| 53 | +service Dapr { |
| 54 | + // Conversate. |
| 55 | + rpc Converse(stream ConversationRequest) returns (stream ConversationResponse); |
| 56 | +} |
| 57 | +
|
| 58 | +// ConversationRequest is the request object for Conversation. |
| 59 | +message ConversationRequest { |
| 60 | + // The name of Conversation component |
| 61 | + string name = 1; |
| 62 | + // Conversation context - the Id of an existing chat room (like in ChatGPT) |
| 63 | + optional string conversationContext = 2; |
| 64 | + // Inputs for the conversation, support multiple input in one time. |
| 65 | + repeated string inputs = 3; |
| 66 | + // Parameters for all custom fields. |
| 67 | + repeated google.protobuf.Any parameters = 4; |
| 68 | +} |
| 69 | +
|
| 70 | +// ConversationResult is the result for one input. |
| 71 | +message ConversationResult { |
| 72 | + // Result for the one conversation input. |
| 73 | + string result = 1; |
| 74 | + // Parameters for all custom fields. |
| 75 | + repeated google.protobuf.Any parameters = 2; |
| 76 | +} |
| 77 | +
|
| 78 | +// ConversationResponse is the response for Conversation. |
| 79 | +message ConversationResponse { |
| 80 | + // Conversation context - the Id of an existing or newly created chat room (like in ChatGPT) |
| 81 | + optional string conversationContext = 1; |
| 82 | +
|
| 83 | + // An array of results. |
| 84 | + repeated ConversationResult outputs = 2; |
| 85 | +} |
| 86 | +``` |
| 87 | + |
| 88 | +## HTTP APIs |
| 89 | + |
| 90 | +The HTTP APIs are same with the gRPC APIs： |
| 91 | + |
| 92 | +`POST /v1.0/conversation/[component]/converse` -> Conversate |
| 93 | + |
| 94 | +```json |
| 95 | +REQUEST = { |
| 96 | + "conversationContext": "fb512b84-7a1a-4fb4-8bd2-ac7d2ec45984", |
| 97 | + "inputs": ["what is Dapr", "Why use Dapr"], |
| 98 | + "parameters": {}, |
| 99 | +} |
| 100 | +
|
| 101 | +RESPONSE = { |
| 102 | + "conversationContext": "fb512b84-7a1a-4fb4-8bd2-ac7d2ec45984", |
| 103 | + "outputs": { |
| 104 | + { |
| 105 | + "result": "Dapr is distribution application runtime ...", |
| 106 | + "parameters": {}, |
| 107 | + }, |
| 108 | + { |
| 109 | + "result": "Dapr can help developers ...", |
| 110 | + "parameters": {}, |
| 111 | + } |
| 112 | +
|
| 113 | + }, |
| 114 | +} |
| 115 | +``` |
| 116 | + |
| 117 | +> Note: URL will begin with `/v1.0-alpha1` while in preview |
0 commit comments