Prompt management

Prompt chains

Integrating SDK into your code

getPromptChain function will respond with deployed version of the given prompt chain. You can deploy prompt chains directly from the Maxim dashboard.

const promptChain = await maxim.getPromptChain("prompt-chain-id");

For a prompt chain with specific deployment variables

For building query to get prompt chain with specific deployment variables, you can use QueryBuilder.

const promptChain = await maxim.getPromptChain("prompt-chain-id",
					new QueryBuilder()
						.and()
						.deploymentVar("Environment", "prod")
						.build()});

Adding multiple queries

const promptChain = await maxim.getPromptChain(
	"prompt-chain-id",
	new QueryBuilder()
	    .and()
		.deploymentVar("Environment", "prod")
		.deploymentVar("CustomerId", "123")
		.build(),
);

Querying Prompts

Sometimes you have usescases where you need to fetch multiple deployed prompt chains at once using a single query. For example, you might want to fetch all prompts for a specific customer or specific workflow. You can use getPromptChains function for this purpose.

You will need to query using at-least one deploymentVar as a filter. Hence you will need to deploy prompt chain versions before querying them.

Query deployed prompt chains using folder

To get all prompts from a folder, you can use getPromptChains function with folderId as a query parameter.

First capture folder id

There are multiple ways to capture folder id. You can use Maxim dashboard to get folder id.

Right click/click on three dots on the folder you want to get id for.
Select Edit Folder option.
You will see folder id in the form.

const folder = await maxim.getFolderById("folder-id");

To get folders using tags attached to the folder.

const folders = await maxim.getFolders(new QueryBuilder().and().tag("CustomerId", "123").build());

All the rules of prompt chain matching algorithm apply here. You can use same overriding techniques as explained above.

Get all deployed prompts from a folder

const folder = ...
const promptChains = await maxim.getPromptChains(
	new QueryBuilder()
	.and()
	.folder(folder.id)
	.deploymentVar("Environment", "prod")
	.build(),
);

You can filter prompt chains based on deploymentVars.

const folder = ...
const promptChains = await maxim.getPromptChains(
		new QueryBuilder()
		.and()
		.folder(folder.id)
		.deploymentVar("Environment", "prod")
		.build());

You have to pass at-least one filter along with folder().

Prompt chain Structure

export type PromptChain = {
	promptChainId: string;
	version: number;
	versionId: string;
	nodes: ({ order: number } & PromptNode)[];
};
// Prompt node
export type PromptNode = {
	prompt: Prompt;
};
// Prompt
export type Prompt = {
	promptId: string;
	version: number;
	versionId: string;
	messages: { role: string; content: string | CompletionRequestContent[] }[];
	modelParameters: { [key: string]: any };
	model: string;
	tags: PromptTags;
};

Folder Structure

export type Folder = {
	id: string;
	name: string;
	parentFolderId?: string;
	tags: { [key: string]: string };
};

Using your own cache for prompts

Maxim SDK uses in-memory caching by default. You can use your own caching implementation by passing a custom cache object to the SDK. This allows you to remove complete dependency on our backend.

Interface for custom cache

export interface MaximCache {
	getAllKeys(): Promise<string[]>;
	get(key: string): Promise<string | null>;
	set(key: string, value: string): Promise<void>;
	delete(key: string): Promise<void>;
}

You will have to pass this custom cache object to the SDK while initializing it.

const maxim = new Maxim({ apiKey: "api-key", cache: new CustomCache() });

Example

Here is the default in-memory cache implementation used by the SDK.

export class MaximInMemoryCache implements MaximCache {
	private cache: Map<string, string> = new Map();
 
	getAllKeys(): Promise<string[]> {
		return Promise.resolve(Array.from(this.cache.keys()));
	}
 
	get(key: string): Promise<string | null> {
		return Promise.resolve(this.cache.get(key) || null);
	}
	set(key: string, value: string): Promise<void> {
		this.cache.set(key, value);
		return Promise.resolve();
	}
}

Matching algorithm

Before going into the details of how to use the SDK, let's understand how the matching algorithm works. Maxim SDK uses best matching entity algorithm.

Let's assume that, you have asked for a prompt chain with deployment var env as prod, customerId as "123" and a tag, tenantId as 456 for promptId - "abc".
SDK will first try to find a prompt chain matching all conditions.
If we don't find any matching entity, we enforce only deploymentVar conditions (you can override this behaviour, as explained in the next section) and match as many tags as possible.
If we still don't find any prompt chain, we check for a prompt chain version marked as fallback.
If we still don't find any prompt chain, we return null.

Overriding fallback algorithm

You can override fallback algorithm by calling .exactMatch() on QueryBuilder object. That will enforce all conditions to be matched.

const promptChain = await maxim.getPromptChain(
	"prompt-chain-id",
	new QueryBuilder()
	    .and()
		.deploymentVar("Environment", "prod")
		.exactMatch()
		.build(),
);

You can override fallback algorithm at each variable level. The third optional parameter in deploymentVar & tag function is enforce. If you pass true to this parameter, it will enforce exact match for that variable.

const promptChain = await maxim.getPromptChain(
    "prompt-id",
    new QueryBuilder()
        .and()
        .deploymentVar("Environment", "prod")
        .build()
);

Prompts

Datasets

Prompt chains

On this page