How to chain prompts in teams-ai version 1.0.0 (MS Teams AI Bot app)?

[donatas-xyz]

I'm raising this question here as suggested by @corinagum.

In @microsoft/teams-ai: ^0.1.0-preview.2 I used to use AI<ApplicationTurnState>.chain(context: TurnContext, state: ApplicationTurnState, prompt?: string | PromptTemplate, options?: Partial<AIOptions<ApplicationTurnState>>): Promise<...> Chains into another prompt and executes the plan that is returned..

Example:

await app.ai.chain(context, state, 'generateTenancyAgreement');

However in the new @microsoft/teams-ai: ^1.0.0 I cannot for the life of me find anything like that? The closest I've got is ActionPlanner<TurnState<DefaultConversationState, DefaultUserState, DefaultTempState>>.completePrompt<string>(context: TurnContext, memory: Memory, prompt: string | PromptTemplate, validator?: PromptResponseValidator<string>): Promise<...> Completes a prompt using an optional validator..

Example:

const response = await planner.completePrompt(context, state, 'generateTenancyAgreement');

I've looked at all the updated examples and couldn't see anything in regards of this functionality?

I think I've managed to address all the other breaking changes but this one.

As you can see from the example below, the generateTenancyAgreement prompt is chained onto chat prompt like so:
general chat starts > "generate tenancy agreement" > bot requests property address following instructions inside the chat prompt > generateTenancyAgreement.ts is executed following chat prompt logic still > await app.ai.chain(context, state, 'generateTenancyAgreement'); is used to chain src\prompt\generateTenancyAgreement with it's own parameters and prompt instructions > src\prompt\generateTenancyAgreement\skprompt.txt are being followed and expected result returned.

How would you achieve the same in version 1.0.0?

EDIT: It would seem I can use planner.completePrompt(context, state, 'generateTenancyAgreement') instead of chain(). It may be that the bug #1130 was throwing me off massively in my efforts or something changed slightly since teams-ai 1.3.

Additional information

.chain() example

src\prompts\chat

config.json

{
  "schema": 1,
  "description": "AI Bot",
  "type": "completion",
  "completion": {
    "max_tokens": 2000,
    "temperature": 0.8,
    "top_p": 0.0,
    "presence_penalty": 0.0,
    "frequency_penalty": 0.0,
    "stop_sequences": [
      "Person:",
      "Bot:"
    ]
  }
}

skprompt.txt (simplified)

#### START PROMPT INSTRUCTIONS

## Prompt grounding
- You are an AI assistant that helps people find information. 

#### END PROMPT INSTRUCTIONS

#### START ACTION INSTRUCTIONS 

Only the following actions and functions are supported:
- {"type":"DO","action":"generateTenancyAgreement","entities":{"operation": "generate", "address": "<propertyAddress>"}}

When to use actions and functions:
- Use `generateTenancyAgreement` to generate an assured shorthold tenancy agreement when the Person asks to create a tenancy agreement 
- Always try to use actions or funtions first, before returning model's response 

What to do after action and functions have been executed: 
- Return the action results to the person. 
- If the chat doesn't match the action then just use standard chat. 
- The Assistant must return the following JSON structure if it matches an action:
{"type":"plan","commands":[{"type":"DO","action":"<name>","entities":{"<name>":<value>}},{"type":"SAY","response":"<response>"}]}

#### END ACTION INSTRUCTIONS 

#### START ACTION EXAMPLES

Person: generate a tenancy agreement 
Bot: What is the property address? 
Person: {{$input}} 
Bot: DO generateTenancyAgreement operation="generate" address="{{$input}}" THEN SAY Can I help you with anything else? 

#### END ACTION EXAMPLES

{{$history}}
Person: {{$input}}
Bot: 

generateTenancyAgreement.ts

export function generateTenancyAgreement(app: Application<ApplicationTurnState>): void 
{
    app.ai.action("generateTenancyAgreement", async (context: TurnContext, state: ApplicationTurnState, data: ITenancyAgreement) => 
    {
        const action = data.operation!.toLowerCase();
        const address = data.address!;
        state.temp.value.propertyAddress = address;
      
        switch (action) 
        {
            case 'generate':
                await context.sendActivity(`&#8987; Generating a Tenancy Agreement for **'${address}'**...`);
                await app.ai.chain(context, state, 'generateTenancyAgreement');
                return false;
            default:
                await context.sendActivity(`Oops, my developer hasn't implemented the ${action} path of 'generateTenancyAgreement' yet!`);
                return true;
        }
    });
}

src\prompts\generateTenancyAgreement

config.json

{
	"schema": 1,
	"description": "Generate Tenancy Agreement.",
	"type": "completion",
	"completion": {
		"max_tokens": 4000,
		"temperature": 0.4,
		"top_p": 0.0,
		"presence_penalty": 0.0,
		"frequency_penalty": 0.0,
		"stop_sequences": [
			"Person:",
			"Bot:"
		]
	},
	"input": {
		"parameters": [
			{
				"name": "history",
				"description": "The history of the conversation.",
				"defaultValue": ""
			},
			{
				"name": "address",
				"description": "Property address.",
				"defaultValue": ""
			}
		]
	}
}

skprompt.txt (simplified)

You are a Legal Advisor at ... 
Your name is Bot. 
Your job is to provide guidance on tenancy agreements. 

You have been asked by the Person to generate Assured Shorthold Tenancy Agreement for your client to read and sign.

ASSUMPTIONS:
Use following **information** to generate the agreement: 
- The agreement is for a residential property. 
- The agreement is for a fixed term of 12 months. 
- The property address is {{$propertyAddress}}

STEPS:
1. Use the **information** provided and the **template** below to generate the agreement: 
	1.1.1. Paragraph headers inside the "agreementBody" node should start a numbered list 

{agreementBodyJsonString-Template}

2. Using the information inside the **template** generate a human readable **text document** ready to be printed out 
	2.1. Format the "header" nodes as paragraph headers 
		2.1.1. Paragraph headers should be in bold text 
	3. **Display** the **text document** 

Let's do it in steps 

RULES: 
- Show **only** the information in the **text document** and **nothing** else 

Bot: **text document**

• StackOverflow query
• GitHub query

[corinagum]

Hi @donatas-xyz, thank you for posting! We have indeed deprecated the chain() feature, but chaining actions should still be possible using app.ai.doAction

Could you share your actions.json file? In case the file hasn't been created yet, I'm wondering if that or some formatting issue may be interfering with the bot's functionality.

[corinagum]

Hi @donatas-xyz, my sincere apologies! You are quite correct and I have gotten my wires crossed. Let me circle back and get back to you asap.

[corinagum]

@donatas-xyz, Thanks for your patience! To re-iterate, prompt chaining was removed from the SDK before our release to GA. Pre-GA, everything was considered experimental and not guaranteed to remain in the codebase. At the moment, there's no direct replacement like what you're discussing above in your clarification - the way we have now to affect chaining is through actions, as described in my original reply.

To get more insight on why chain() was removed, @Stevenic would be able to provide more information. Steve, if you see this, could you provide details on the decision for removing chain()?

Otherwise, if there is enough customer interest in bringing back chain(), it's certainly something we can look into adding to our roadmap. As of right now though, there's no plan for a direct replacement.

If there are other developers who are interested in prompt chaining like chain() used to do, let's please use this discussion item as a sounding board for people to voice their thoughts. If there is enough engagement, I can return to my team to discuss the option of bringing it back.

Sorry this isn't the response you are looking for, but I hope that clarifies things!

[lifan0127]

Hi I am late in the game so I don't have experience with app.ai.chain. I have used the planner.completePrompt approach as mentioned by @donatas-xyz to achieve similar functionality, but it requires me to handle the response manually with doAction etc. It seems that app.ai.chain is more straightforward /elegant.

By the way, @corinagum I believe we are missing an equivalent do_action method for the Python implementation.

[kevmcdonk]

I was using Chain in the sample for Graph SSO and found it was a great way for using RAG pattern and then returning it with a prompt. Is there really no way of doing this now?

https://github.com/kevmcdonk/TeamsAI-GraphSample/blob/main/index.ts

[corinagum]

Hi @donatas-xyz, @lifan0127, @kevmcdonk,

Because there's been a lot of discussion about chain() since it got deprecated, I brought it up with my team. Currently there is no plan to bring chain back since doAction is available, however, if any of you are still keen on the feature, please feel free to file 1 (one) feature request to bring chain back for prompt kickoff.

I can't guarantee it will be brought back, but the more of you who +1 the feature request, the more weight is given for consideration when planning our roadmap :) If you can spread the word to get upvotes/reactions that of course helps too.

@lifan0127, I saw your comment about do_action for Python and filed a work item for that. Thanks for the heads up.