LLM - LLMAssistantBuilder

LLM.LLMAssistantBuilder

Construct a multi-modal LLM agent from a purpose, prompt, functions and streams.

Type parameters

Name	Type
Inputs	extends Record = Record
Functions	extends Record = Record
Outputs	extends Record = Record
Samples	extends VariantType<{ message: LLMMessage } & { [K in keyof Functions]: LLMFunctionCall }> = VariantType<{ message: LLMMessage } & { [K in keyof Functions]: LLMFunctionCall }>

Hierarchy

• Builder

  ↳ LLMAssistantBuilder

LLM

aggregate

▸ aggregate(function_name, config):

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMAggregateFunction }, Outputs, Samples>

Allow the agent provide definition of an aggregation to perform on a stream, if added to a layout the values will be shown in a chat message.

Type parameters

Name	Type
Name	extends string
R	extends DictType

Parameters

Name	Type	Description
function_name	Name extends keyof Functions ? never : Name	The unique identifier for the function.
config	Object	A configuration object providing options for the aggregation operation.
config.function_description?	string	The description of the function. *
config.silent?	boolean	False if the function should return the value to the agent. *
config.value	(inputs: Inputs) => Variable	The Variable to be aggregated. *
config.value_description?	AbstractTypeBuilder | (builder: TypeBuilder) => AbstractTypeBuilder	The optional description of the entries in the collection. *

Returns

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMAggregateFunction }, Outputs, Samples>

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .aggregate(
      "aggregate_sales",
      {
         value: (inputs) => inputs.sales,
      }
   )

---

chart

▸ chart(function_name, config):

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMChartFunction }, Outputs, Samples>

Allow the agent to create a chart from a stream, if added to a layout the chart will be shown in a chat message.

Type parameters

Name	Type
Name	extends string
T	extends DictType

Parameters

Name	Type	Description
function_name	Name extends keyof Functions ? never : Name	The unique identifier for the function.
config	Object	A configuration object providing options for the chart operation.
config.function_description?	string	The description of the function. *
config.value	Stream	The Stream to visualize. *
config.value_description?	AbstractTypeBuilder | (builder: TypeBuilder) => AbstractTypeBuilder	The optional description of the Stream to visualize. *

Returns

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMChartFunction }, Outputs, Samples>

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .chart(
       "chart_sales_data",
       {
         value: sales_data.outputStream(),
       }
   )

---

example

▸ example(values):

LLMAssistantBuilder

Add a chain of example messages, and functions calls to inform the agent.

Parameters

Name	Type	Description
values	(builder: LLMTrainingSampleBuilder<Functions, VariantType<{ message: LLMMessage } & { [K in string | number | symbol]: LLMFunctionCall }>>) => LLMTrainingSampleBuilder<Functions, VariantType<{ message: LLMMessage } & { [K in string | number | symbol]: LLMFunctionCall }>>	The builder defining the examples.

Returns

LLMAssistantBuilder

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .chart(
       "chart_sales_data",
       `Read, aggregate and chart the sales data`,
       sales_data.outputStream(),
   )
   .example(builder => builder
       .message('user', "Can you show me the revenue over time per item, based on a price of $5.5?")
       .call('write_inputs', builder => builder.args("Set the price.", 5.5))
       .call('read_and_aggregate_results_data', builder => builder
         .args(
             "A chart showing the total revenue per item over time.",
             {
                 name: "Total revenue per item over time",
                 mark: { kind: "line" },
                 x: { field: 'date', title: "Date", type: 'temporal' },
                 y: { field: 'amount', title: "Revenue", type: 'quantitative', aggregate: 'sum' },
                 color: { field: 'item', title: "Item", type: 'nominal' }
             })
       )
       .message('assistant', "The chart of total revenue for a price of $5.5 was generated successfully.")
   )

---

expression

▸ expression(function_name, config):

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMExpressionFunction }, Outputs, Samples>

Allow the agent provide arguments to a function that evaluates a user defined expression, if added to a layout the values will be shown in a chat message.

Type parameters

Name	Type
Name	extends string
I	extends EastType
R	extends EastType

Parameters

Name	Type	Description
function_name	Name extends keyof Functions ? never : Name	The unique identifier for the function.
config	Object	A configuration object providing options for the expression operation.
config.function_args	AbstractTypeBuilder | (builder: TypeBuilder) => AbstractTypeBuilder	The StructType function arguments that must be passed in by the agent. *
config.function_description?	string	The description of the function. *
config.return_description?	AbstractTypeBuilder | (builder: TypeBuilder) => AbstractTypeBuilder	The optional description of the return value. *
config.return_error?	Object	The predicate to check for an error, and message returned to the agent based on the arguments, return value and inputs. *
config.return_error.is_error	(args: Variable, ret: Variable, inputs: Inputs) => EastFunction	the EastFunction with a predicate indicating if an has occurred
config.return_error.message	(args: Variable, ret: Variable, inputs: Inputs) => EastFunction	the EastFunction with an error message
config.return_value	(args: Variable, inputs: Inputs) => EastFunction	The EastFunction value returned to the agent. *
config.silent?	boolean	False if the function should return the value to the agent, defaults to true. *

Returns

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMExpressionFunction }, Outputs, Samples>

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .expression(
       "filter_sales",
       {
           function_args: (builder) => builder
               .describe("The filter parameters")
               .struct(builder => builder
                   .field('start', builder => builder.describe('sales with dates earlier than this will be removed').datetime())
                   .field('end', builder => builder.describe('sales with dates later than this will be removed').datetime())
           ),
           return_value: (args, inputs) => Filter(
               inputs.sales,
               (value) => And(
                   GreaterEqual(GetField(value, 'date'), GetField(args, 'start')),
                   LessEqual(GetField(value, 'date'), GetField(args, 'end'))
               )
           ),
       }
   )

---

input

▸ input(config):

LLMAssistantBuilder<Inputs & { [K in string]: Variable }, Functions, Outputs, Samples>

Add an additional named input

Stream to the LLMBuilder.

Type parameters

Name	Type
Name	extends string
I	extends EastType

Parameters

Name	Type	Description
config	Object	the input stream and the resulting variable name
config.name	Name extends "input" | keyof Inputs ? never : Name	The name of the input. *
config.stream	Stream	The Stream to input. *

Returns

LLMAssistantBuilder<Inputs & { [K in string]: Variable }, Functions, Outputs, Samples>

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .read(
        "get_the_price",
        "Get the current value of the price",
        (inputs) => inputs.price,
    )

---

output

▸ output(function_name, config):

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMOutputFunction }, Outputs & { [K in string]: Stream }, Samples>

Allow the agent to produce a stream output, if added to a layout the values will be shown in a chat message.

Type parameters

Name	Type
Name	extends string
I	extends EastType

Parameters

Name	Type	Description
function_name	Name extends keyof Functions ? never : Name	The unique identifier for the function.
config	Object	A configuration object providing options for the output operation.
config.function_description?	string	The description of the function. *
config.return_error?	Object	The predicate to check for an error, and message returned to the agent based on the arguments, and inputs. *
config.return_error.is_error	(args: Variable, inputs: Inputs) => EastFunction	the EastFunction with a predicate indicating if an has occurred
config.return_error.message	(args: Variable, inputs: Inputs) => EastFunction	the EastFunction with an error message
config.value	AbstractTypeBuilder | (builder: TypeBuilder) => AbstractTypeBuilder	The TypeBuilder describing the output value. *

Returns

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMOutputFunction }, Outputs & { [K in string]: Stream }, Samples>

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .output(
      "change_the_price",
      {
          value: (builder) => builder.describe("The new price").float(),
      }
   )

---

read

▸ read(function_name, config):

LLMAssistantBuilder<Inputs & { [K in string]: Variable }, Functions & { [K in string]: LLMValueFunction }, Outputs, Samples>

Allow the agent to read a

Stream value, if added to a layout the values will be shown in a chat message.

Type parameters

Name	Type
Name	extends string
T	extends EastType

Parameters

Name	Type	Description
function_name	Name extends keyof Functions ? never : Name	The unique identifier for the function.
config	Object	A configuration object providing options for the read operation.
config.function_description?	string	The description of the function. *
config.return_description?	AbstractTypeBuilder | (builder: TypeBuilder) => AbstractTypeBuilder	The optional description of the stream value. *
config.return_value	(inputs: Inputs) => Variable	The value returned to the assistant, Remarks this is not an Expression *
config.silent?	boolean	False if the function should return the value to the agent. *

Returns

LLMAssistantBuilder<Inputs & { [K in string]: Variable }, Functions & { [K in string]: LLMValueFunction }, Outputs, Samples>

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .read(
      "get_the_price",
      {
         return_value: (inputs) => inputs.price, 
      }
   )

---

write

▸ write(function_name, config):

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMWriteFunction }, Outputs, Samples>

Allow the agent to write a value to an existing writable stream, if added to a layout the values will be shown in a chat message.

Type parameters

Name	Type
Name	extends string
T	extends EastType

Parameters

Name	Type	Description
function_name	Name extends keyof Functions ? never : Name	The unique identifier for the function.
config	Object	A configuration object providing options for the write operation.
config.function_description?	string	The description of the function. *
config.return_error?	Object	The predicate to check for an error, and message returned to the agent based on the arguments, and inputs. *
config.return_error.is_error	(args: Variable, inputs: Inputs) => EastFunction	the EastFunction with a predicate indicating if an has occurred
config.return_error.message	(args: Variable, inputs: Inputs) => EastFunction	the EastFunction with an error message
config.value	Stream	The stream containing the value *
config.value_description?	AbstractTypeBuilder | (builder: TypeBuilder) => AbstractTypeBuilder	The optional description of the stream. *

Returns

LLMAssistantBuilder<Inputs, Functions & { [K in string]: LLMWriteFunction }, Outputs, Samples>

a new

LLMBuilder

Example

 const price = new SourceBuilder("price")
   .value({ value: 5 });

 const message = new SourceBuilder("message",)
   .writeable(AssistantInputStateType);

 const sales = new SourceBuilder("sales")
   .writeable(DictType(
     StringType,
     StructType({
       date: DateTimeType,
       item: StringType,
       qty: IntegerType,
       inventory: IntegerType,
       price: FloatType,
       amount: FloatType,
     }))
   )

 const model = new SourceBuilder("model",)
   .value({ value: variant('gpt-3.5-turbo-1106', null), type: LLMModelType });
 
 const assistant = new LLMBuilder("assistant")
   .input({ name: "sales", stream: sales.outputStream() })
   .input({ name: "message", stream: message.outputStream() })
   .input({ name: "model", stream: model.outputStream() })
   .input({ name: "price", stream: price.outputStream() })
   .assistant({
       api_key: "...",
       purpose: "You are a business analyst assistant",
       prompt: (inputs) => inputs.message,
       model: (inputs) => inputs.model,
   })
   .write(
      "change_the_price",
      {
         value: price.outputStream(),
      }
   )

Source

outputStreams

▸ outputStreams(): Outputs

Return the

Stream containing the output of the agent.

Returns

Outputs