mirror of
https://github.com/farcasclaudiu/Flowise.git
synced 2026-06-29 01:01:11 +03:00
Feature/seq agents (#2798)
* update build functions * sequential agents * update langchain to 0.2, added sequential agent nodes * add marketplace templates * update howto wordings * Merge branch 'main' into feature/Seq-Agents # Conflicts: # pnpm-lock.yaml * update deprecated functions and add new sequential nodes * add marketplace templates * update marketplace templates, add structured output to llm node * add multi agents template * update llm node with bindmodels * update cypress version * update templates sticky note wordings * update tool node to include human in loop action * update structured outputs error from models * update cohere package to resolve google genai pipeThrough bug * update mistral package version, added message reconstruction before invoke seq agent * add HITL to agent * update state messages restructuring * update load and split methods for s3 directory
This commit is contained in:
Henry Heng
@@ -0,0 +1,886 @@
|
||||
import { flatten, uniq } from 'lodash'
|
||||
import { DataSource } from 'typeorm'
|
||||
import { RunnableSequence, RunnablePassthrough, RunnableConfig } from '@langchain/core/runnables'
|
||||
import { ChatPromptTemplate, MessagesPlaceholder, HumanMessagePromptTemplate, BaseMessagePromptTemplateLike } from '@langchain/core/prompts'
|
||||
import { BaseChatModel } from '@langchain/core/language_models/chat_models'
|
||||
import { AIMessage, AIMessageChunk, BaseMessage, HumanMessage, ToolMessage } from '@langchain/core/messages'
|
||||
import { formatToOpenAIToolMessages } from 'langchain/agents/format_scratchpad/openai_tools'
|
||||
import { type ToolsAgentStep } from 'langchain/agents/openai/output_parser'
|
||||
import { StringOutputParser } from '@langchain/core/output_parsers'
|
||||
import {
|
||||
INode,
|
||||
INodeData,
|
||||
INodeParams,
|
||||
ISeqAgentsState,
|
||||
ICommonObject,
|
||||
MessageContentImageUrl,
|
||||
INodeOutputsValue,
|
||||
ISeqAgentNode,
|
||||
IDatabaseEntity,
|
||||
IUsedTool,
|
||||
IDocument
|
||||
} from '../../../src/Interface'
|
||||
import { ToolCallingAgentOutputParser, AgentExecutor, SOURCE_DOCUMENTS_PREFIX } from '../../../src/agents'
|
||||
import { getInputVariables, getVars, handleEscapeCharacters, prepareSandboxVars } from '../../../src/utils'
|
||||
import {
|
||||
customGet,
|
||||
getVM,
|
||||
processImageMessage,
|
||||
transformObjectPropertyToFunction,
|
||||
restructureMessages,
|
||||
MessagesState,
|
||||
RunnableCallable
|
||||
} from '../commonUtils'
|
||||
import { END, StateGraph } from '@langchain/langgraph'
|
||||
import { StructuredTool } from '@langchain/core/tools'
|
||||
|
||||
const defaultApprovalPrompt = `You are about to execute tool: {tools}. Ask if user want to proceed`
|
||||
const examplePrompt = 'You are a research assistant who can search for up-to-date info using search engine.'
|
||||
const customOutputFuncDesc = `This is only applicable when you have a custom State at the START node. After agent execution, you might want to update the State values`
|
||||
const howToUseCode = `
|
||||
1. Return the key value JSON object. For example: if you have the following State:
|
||||
\`\`\`json
|
||||
{
|
||||
"user": null
|
||||
}
|
||||
\`\`\`
|
||||
|
||||
You can update the "user" value by returning the following:
|
||||
\`\`\`js
|
||||
return {
|
||||
"user": "john doe"
|
||||
}
|
||||
\`\`\`
|
||||
|
||||
2. If you want to use the agent's output as the value to update state, it is available as \`$flow.output\` with the following structure:
|
||||
\`\`\`json
|
||||
{
|
||||
"content": "Hello! How can I assist you today?",
|
||||
"usedTools": [
|
||||
{
|
||||
"tool": "tool-name",
|
||||
"toolInput": "{foo: var}",
|
||||
"toolOutput": "This is the tool's output"
|
||||
}
|
||||
],
|
||||
"sourceDocuments": [
|
||||
{
|
||||
"pageContent": "This is the page content",
|
||||
"metadata": "{foo: var}",
|
||||
}
|
||||
],
|
||||
}
|
||||
\`\`\`
|
||||
|
||||
For example, if the \`toolOutput\` is the value you want to update the state with, you can return the following:
|
||||
\`\`\`js
|
||||
return {
|
||||
"user": $flow.output.usedTools[0].toolOutput
|
||||
}
|
||||
\`\`\`
|
||||
|
||||
3. You can also get default flow config, including the current "state":
|
||||
- \`$flow.sessionId\`
|
||||
- \`$flow.chatId\`
|
||||
- \`$flow.chatflowId\`
|
||||
- \`$flow.input\`
|
||||
- \`$flow.state\`
|
||||
|
||||
4. You can get custom variables: \`$vars.<variable-name>\`
|
||||
|
||||
`
|
||||
const howToUse = `
|
||||
1. Key and value pair to be updated. For example: if you have the following State:
|
||||
| Key | Operation | Default Value |
|
||||
|-----------|---------------|-------------------|
|
||||
| user | Replace | |
|
||||
|
||||
You can update the "user" value with the following:
|
||||
| Key | Value |
|
||||
|-----------|-----------|
|
||||
| user | john doe |
|
||||
|
||||
2. If you want to use the agent's output as the value to update state, it is available as available as \`$flow.output\` with the following structure:
|
||||
\`\`\`json
|
||||
{
|
||||
"output": "Hello! How can I assist you today?",
|
||||
"usedTools": [
|
||||
{
|
||||
"tool": "tool-name",
|
||||
"toolInput": "{foo: var}",
|
||||
"toolOutput": "This is the tool's output"
|
||||
}
|
||||
],
|
||||
"sourceDocuments": [
|
||||
{
|
||||
"pageContent": "This is the page content",
|
||||
"metadata": "{foo: var}",
|
||||
}
|
||||
],
|
||||
}
|
||||
\`\`\`
|
||||
|
||||
For example, if the \`toolOutput\` is the value you want to update the state with, you can do the following:
|
||||
| Key | Value |
|
||||
|-----------|-------------------------------------------|
|
||||
| user | \`$flow.output.usedTools[0].toolOutput\` |
|
||||
|
||||
3. You can get default flow config, including the current "state":
|
||||
- \`$flow.sessionId\`
|
||||
- \`$flow.chatId\`
|
||||
- \`$flow.chatflowId\`
|
||||
- \`$flow.input\`
|
||||
- \`$flow.state\`
|
||||
|
||||
4. You can get custom variables: \`$vars.<variable-name>\`
|
||||
|
||||
`
|
||||
const defaultFunc = `const result = $flow.output;
|
||||
|
||||
/* Suppose we have a custom State schema like this:
|
||||
* {
|
||||
aggregate: {
|
||||
value: (x, y) => x.concat(y),
|
||||
default: () => []
|
||||
}
|
||||
}
|
||||
*/
|
||||
|
||||
return {
|
||||
aggregate: [result.content]
|
||||
};`
|
||||
const TAB_IDENTIFIER = 'selectedUpdateStateMemoryTab'
|
||||
|
||||
class Agent_SeqAgents implements INode {
|
||||
label: string
|
||||
name: string
|
||||
version: number
|
||||
description: string
|
||||
type: string
|
||||
icon: string
|
||||
category: string
|
||||
baseClasses: string[]
|
||||
inputs?: INodeParams[]
|
||||
badge?: string
|
||||
outputs: INodeOutputsValue[]
|
||||
|
||||
constructor() {
|
||||
this.label = 'Agent'
|
||||
this.name = 'seqAgent'
|
||||
this.version = 1.0
|
||||
this.type = 'Agent'
|
||||
this.icon = 'seqAgent.png'
|
||||
this.category = 'Sequential Agents'
|
||||
this.description = 'Agent that can execute tools'
|
||||
this.baseClasses = [this.type]
|
||||
this.inputs = [
|
||||
{
|
||||
label: 'Agent Name',
|
||||
name: 'agentName',
|
||||
type: 'string',
|
||||
placeholder: 'Agent'
|
||||
},
|
||||
{
|
||||
label: 'System Prompt',
|
||||
name: 'systemMessagePrompt',
|
||||
type: 'string',
|
||||
rows: 4,
|
||||
optional: true,
|
||||
default: examplePrompt
|
||||
},
|
||||
{
|
||||
label: 'Human Prompt',
|
||||
name: 'humanMessagePrompt',
|
||||
type: 'string',
|
||||
description: 'This prompt will be added at the end of the messages as human message',
|
||||
rows: 4,
|
||||
optional: true,
|
||||
additionalParams: true
|
||||
},
|
||||
{
|
||||
label: 'Tools',
|
||||
name: 'tools',
|
||||
type: 'Tool',
|
||||
list: true
|
||||
},
|
||||
{
|
||||
label: 'Start | Agent | LLM | Tool Node',
|
||||
name: 'sequentialNode',
|
||||
type: 'Start | Agent | LLMNode | ToolNode',
|
||||
list: true
|
||||
},
|
||||
{
|
||||
label: 'Chat Model',
|
||||
name: 'model',
|
||||
type: 'BaseChatModel',
|
||||
optional: true,
|
||||
description: `Overwrite model to be used for this agent`
|
||||
},
|
||||
{
|
||||
label: 'Require Approval',
|
||||
name: 'interrupt',
|
||||
description: 'Require approval before executing tools. Will proceed when tools are not called',
|
||||
type: 'boolean',
|
||||
optional: true
|
||||
},
|
||||
{
|
||||
label: 'Format Prompt Values',
|
||||
name: 'promptValues',
|
||||
description: 'Assign values to the prompt variables. You can also use $flow.state.<variable-name> to get the state value',
|
||||
type: 'json',
|
||||
optional: true,
|
||||
acceptVariable: true,
|
||||
list: true
|
||||
},
|
||||
{
|
||||
label: 'Approval Prompt',
|
||||
name: 'approvalPrompt',
|
||||
description: 'Prompt for approval. Only applicable if "Require Approval" is enabled',
|
||||
type: 'string',
|
||||
default: defaultApprovalPrompt,
|
||||
rows: 4,
|
||||
optional: true,
|
||||
additionalParams: true
|
||||
},
|
||||
{
|
||||
label: 'Approve Button Text',
|
||||
name: 'approveButtonText',
|
||||
description: 'Text for approve button. Only applicable if "Require Approval" is enabled',
|
||||
type: 'string',
|
||||
default: 'Yes',
|
||||
optional: true,
|
||||
additionalParams: true
|
||||
},
|
||||
{
|
||||
label: 'Reject Button Text',
|
||||
name: 'rejectButtonText',
|
||||
description: 'Text for reject button. Only applicable if "Require Approval" is enabled',
|
||||
type: 'string',
|
||||
default: 'No',
|
||||
optional: true,
|
||||
additionalParams: true
|
||||
},
|
||||
{
|
||||
label: 'Update State',
|
||||
name: 'updateStateMemory',
|
||||
type: 'tabs',
|
||||
tabIdentifier: TAB_IDENTIFIER,
|
||||
additionalParams: true,
|
||||
default: 'updateStateMemoryUI',
|
||||
tabs: [
|
||||
{
|
||||
label: 'Update State (Table)',
|
||||
name: 'updateStateMemoryUI',
|
||||
type: 'datagrid',
|
||||
hint: {
|
||||
label: 'How to use',
|
||||
value: howToUse
|
||||
},
|
||||
description: customOutputFuncDesc,
|
||||
datagrid: [
|
||||
{
|
||||
field: 'key',
|
||||
headerName: 'Key',
|
||||
type: 'asyncSingleSelect',
|
||||
loadMethod: 'loadStateKeys',
|
||||
flex: 0.5,
|
||||
editable: true
|
||||
},
|
||||
{
|
||||
field: 'value',
|
||||
headerName: 'Value',
|
||||
type: 'freeSolo',
|
||||
valueOptions: [
|
||||
{
|
||||
label: 'Agent Output (string)',
|
||||
value: '$flow.output.content'
|
||||
},
|
||||
{
|
||||
label: `Used Tools (array)`,
|
||||
value: '$flow.output.usedTools'
|
||||
},
|
||||
{
|
||||
label: `First Tool Output (string)`,
|
||||
value: '$flow.output.usedTools[0].toolOutput'
|
||||
},
|
||||
{
|
||||
label: 'Source Documents (array)',
|
||||
value: '$flow.output.sourceDocuments'
|
||||
},
|
||||
{
|
||||
label: `Global variable (string)`,
|
||||
value: '$vars.<variable-name>'
|
||||
},
|
||||
{
|
||||
label: 'Input Question (string)',
|
||||
value: '$flow.input'
|
||||
},
|
||||
{
|
||||
label: 'Session Id (string)',
|
||||
value: '$flow.sessionId'
|
||||
},
|
||||
{
|
||||
label: 'Chat Id (string)',
|
||||
value: '$flow.chatId'
|
||||
},
|
||||
{
|
||||
label: 'Chatflow Id (string)',
|
||||
value: '$flow.chatflowId'
|
||||
}
|
||||
],
|
||||
editable: true,
|
||||
flex: 1
|
||||
}
|
||||
],
|
||||
optional: true,
|
||||
additionalParams: true
|
||||
},
|
||||
{
|
||||
label: 'Update State (Code)',
|
||||
name: 'updateStateMemoryCode',
|
||||
type: 'code',
|
||||
hint: {
|
||||
label: 'How to use',
|
||||
value: howToUseCode
|
||||
},
|
||||
description: `${customOutputFuncDesc}. Must return an object representing the state`,
|
||||
hideCodeExecute: true,
|
||||
codeExample: defaultFunc,
|
||||
optional: true,
|
||||
additionalParams: true
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
label: 'Max Iterations',
|
||||
name: 'maxIterations',
|
||||
type: 'number',
|
||||
optional: true,
|
||||
additionalParams: true
|
||||
}
|
||||
]
|
||||
}
|
||||
|
||||
async init(nodeData: INodeData, input: string, options: ICommonObject): Promise<any> {
|
||||
let tools = nodeData.inputs?.tools
|
||||
tools = flatten(tools)
|
||||
let agentSystemPrompt = nodeData.inputs?.systemMessagePrompt as string
|
||||
let agentHumanPrompt = nodeData.inputs?.humanMessagePrompt as string
|
||||
const agentLabel = nodeData.inputs?.agentName as string
|
||||
const sequentialNodes = nodeData.inputs?.sequentialNode as ISeqAgentNode[]
|
||||
const maxIterations = nodeData.inputs?.maxIterations as string
|
||||
const model = nodeData.inputs?.model as BaseChatModel
|
||||
const promptValuesStr = nodeData.inputs?.promptValues
|
||||
const output = nodeData.outputs?.output as string
|
||||
const approvalPrompt = nodeData.inputs?.approvalPrompt as string
|
||||
|
||||
if (!agentLabel) throw new Error('Agent name is required!')
|
||||
const agentName = agentLabel.toLowerCase().replace(/\s/g, '_').trim()
|
||||
|
||||
if (!sequentialNodes || !sequentialNodes.length) throw new Error('Agent must have a predecessor!')
|
||||
|
||||
let agentInputVariablesValues: ICommonObject = {}
|
||||
if (promptValuesStr) {
|
||||
try {
|
||||
agentInputVariablesValues = typeof promptValuesStr === 'object' ? promptValuesStr : JSON.parse(promptValuesStr)
|
||||
} catch (exception) {
|
||||
throw new Error("Invalid JSON in the Agent's Prompt Input Values: " + exception)
|
||||
}
|
||||
}
|
||||
agentInputVariablesValues = handleEscapeCharacters(agentInputVariablesValues, true)
|
||||
|
||||
const startLLM = sequentialNodes[0].startLLM
|
||||
const llm = model || startLLM
|
||||
if (nodeData.inputs) nodeData.inputs.model = llm
|
||||
|
||||
const multiModalMessageContent = sequentialNodes[0]?.multiModalMessageContent || (await processImageMessage(llm, nodeData, options))
|
||||
const abortControllerSignal = options.signal as AbortController
|
||||
const agentInputVariables = uniq([...getInputVariables(agentSystemPrompt), ...getInputVariables(agentHumanPrompt)])
|
||||
|
||||
if (!agentInputVariables.every((element) => Object.keys(agentInputVariablesValues).includes(element))) {
|
||||
throw new Error('Agent input variables values are not provided!')
|
||||
}
|
||||
|
||||
const interrupt = nodeData.inputs?.interrupt as boolean
|
||||
|
||||
const toolName = `tool_${nodeData.id}`
|
||||
const toolNode = new ToolNode(tools, nodeData, input, options, toolName, [], { sequentialNodeName: toolName })
|
||||
|
||||
;(toolNode as any).seekPermissionMessage = async (usedTools: IUsedTool[]) => {
|
||||
const prompt = ChatPromptTemplate.fromMessages([['human', approvalPrompt || defaultApprovalPrompt]])
|
||||
const chain = prompt.pipe(startLLM)
|
||||
const response = (await chain.invoke({
|
||||
input: 'Hello there!',
|
||||
tools: JSON.stringify(usedTools)
|
||||
})) as AIMessageChunk
|
||||
return response.content
|
||||
}
|
||||
|
||||
const workerNode = async (state: ISeqAgentsState, config: RunnableConfig) => {
|
||||
return await agentNode(
|
||||
{
|
||||
state,
|
||||
llm,
|
||||
interrupt,
|
||||
agent: await createAgent(
|
||||
agentName,
|
||||
state,
|
||||
llm,
|
||||
interrupt,
|
||||
[...tools],
|
||||
agentSystemPrompt,
|
||||
agentHumanPrompt,
|
||||
multiModalMessageContent,
|
||||
agentInputVariablesValues,
|
||||
maxIterations,
|
||||
{
|
||||
sessionId: options.sessionId,
|
||||
chatId: options.chatId,
|
||||
input
|
||||
}
|
||||
),
|
||||
name: agentName,
|
||||
abortControllerSignal,
|
||||
nodeData,
|
||||
input,
|
||||
options
|
||||
},
|
||||
config
|
||||
)
|
||||
}
|
||||
|
||||
const toolInterrupt = async (
|
||||
graph: StateGraph<any>,
|
||||
nextNodeName?: string,
|
||||
runCondition?: any,
|
||||
conditionalMapping: ICommonObject = {}
|
||||
) => {
|
||||
const routeMessage = async (state: ISeqAgentsState) => {
|
||||
const messages = state.messages as unknown as BaseMessage[]
|
||||
const lastMessage = messages[messages.length - 1] as AIMessage
|
||||
|
||||
if (!lastMessage?.tool_calls?.length) {
|
||||
// if next node is condition node, run the condition
|
||||
if (runCondition) {
|
||||
const returnNodeName = await runCondition(state)
|
||||
return returnNodeName
|
||||
}
|
||||
return nextNodeName || END
|
||||
}
|
||||
return toolName
|
||||
}
|
||||
|
||||
graph.addNode(toolName, toolNode)
|
||||
|
||||
if (nextNodeName) {
|
||||
// @ts-ignore
|
||||
graph.addConditionalEdges(agentName, routeMessage, {
|
||||
[toolName]: toolName,
|
||||
[END]: END,
|
||||
[nextNodeName]: nextNodeName,
|
||||
...conditionalMapping
|
||||
})
|
||||
} else {
|
||||
// @ts-ignore
|
||||
graph.addConditionalEdges(agentName, routeMessage, { [toolName]: toolName, [END]: END, ...conditionalMapping })
|
||||
}
|
||||
|
||||
// @ts-ignore
|
||||
graph.addEdge(toolName, agentName)
|
||||
|
||||
return graph
|
||||
}
|
||||
|
||||
const returnOutput: ISeqAgentNode = {
|
||||
id: nodeData.id,
|
||||
node: workerNode,
|
||||
name: agentName,
|
||||
label: agentLabel,
|
||||
type: 'agent',
|
||||
llm,
|
||||
startLLM,
|
||||
output,
|
||||
predecessorAgents: sequentialNodes,
|
||||
multiModalMessageContent,
|
||||
moderations: sequentialNodes[0]?.moderations,
|
||||
agentInterruptToolNode: interrupt ? toolNode : undefined,
|
||||
agentInterruptToolFunc: interrupt ? toolInterrupt : undefined
|
||||
}
|
||||
|
||||
return returnOutput
|
||||
}
|
||||
}
|
||||
|
||||
async function createAgent(
|
||||
agentName: string,
|
||||
state: ISeqAgentsState,
|
||||
llm: BaseChatModel,
|
||||
interrupt: boolean,
|
||||
tools: any[],
|
||||
systemPrompt: string,
|
||||
humanPrompt: string,
|
||||
multiModalMessageContent: MessageContentImageUrl[],
|
||||
agentInputVariablesValues: ICommonObject,
|
||||
maxIterations?: string,
|
||||
flowObj?: { sessionId?: string; chatId?: string; input?: string }
|
||||
): Promise<any> {
|
||||
if (tools.length && !interrupt) {
|
||||
const promptArrays = [
|
||||
new MessagesPlaceholder('messages'),
|
||||
new MessagesPlaceholder('agent_scratchpad')
|
||||
] as BaseMessagePromptTemplateLike[]
|
||||
if (systemPrompt) promptArrays.unshift(['system', systemPrompt])
|
||||
if (humanPrompt) promptArrays.push(['human', humanPrompt])
|
||||
|
||||
const prompt = ChatPromptTemplate.fromMessages(promptArrays)
|
||||
|
||||
if (multiModalMessageContent.length) {
|
||||
const msg = HumanMessagePromptTemplate.fromTemplate([...multiModalMessageContent])
|
||||
prompt.promptMessages.splice(1, 0, msg)
|
||||
}
|
||||
|
||||
if (llm.bindTools === undefined) {
|
||||
throw new Error(`This agent only compatible with function calling models.`)
|
||||
}
|
||||
const modelWithTools = llm.bindTools(tools)
|
||||
|
||||
let agent
|
||||
|
||||
if (!agentInputVariablesValues || !Object.keys(agentInputVariablesValues).length) {
|
||||
agent = RunnableSequence.from([
|
||||
RunnablePassthrough.assign({
|
||||
//@ts-ignore
|
||||
agent_scratchpad: (input: { steps: ToolsAgentStep[] }) => formatToOpenAIToolMessages(input.steps)
|
||||
}),
|
||||
prompt,
|
||||
modelWithTools,
|
||||
new ToolCallingAgentOutputParser()
|
||||
]).withConfig({
|
||||
metadata: { sequentialNodeName: agentName }
|
||||
})
|
||||
} else {
|
||||
agent = RunnableSequence.from([
|
||||
RunnablePassthrough.assign({
|
||||
//@ts-ignore
|
||||
agent_scratchpad: (input: { steps: ToolsAgentStep[] }) => formatToOpenAIToolMessages(input.steps)
|
||||
}),
|
||||
RunnablePassthrough.assign(transformObjectPropertyToFunction(agentInputVariablesValues, state)),
|
||||
prompt,
|
||||
modelWithTools,
|
||||
new ToolCallingAgentOutputParser()
|
||||
]).withConfig({
|
||||
metadata: { sequentialNodeName: agentName }
|
||||
})
|
||||
}
|
||||
|
||||
const executor = AgentExecutor.fromAgentAndTools({
|
||||
agent,
|
||||
tools,
|
||||
sessionId: flowObj?.sessionId,
|
||||
chatId: flowObj?.chatId,
|
||||
input: flowObj?.input,
|
||||
verbose: process.env.DEBUG === 'true' ? true : false,
|
||||
maxIterations: maxIterations ? parseFloat(maxIterations) : undefined
|
||||
})
|
||||
return executor
|
||||
} else if (tools.length && interrupt) {
|
||||
if (llm.bindTools === undefined) {
|
||||
throw new Error(`Agent Node only compatible with function calling models.`)
|
||||
}
|
||||
// @ts-ignore
|
||||
llm = llm.bindTools(tools)
|
||||
|
||||
const promptArrays = [new MessagesPlaceholder('messages')] as BaseMessagePromptTemplateLike[]
|
||||
if (systemPrompt) promptArrays.unshift(['system', systemPrompt])
|
||||
if (humanPrompt) promptArrays.push(['human', humanPrompt])
|
||||
|
||||
const prompt = ChatPromptTemplate.fromMessages(promptArrays)
|
||||
if (multiModalMessageContent.length) {
|
||||
const msg = HumanMessagePromptTemplate.fromTemplate([...multiModalMessageContent])
|
||||
prompt.promptMessages.splice(1, 0, msg)
|
||||
}
|
||||
|
||||
let agent
|
||||
|
||||
if (!agentInputVariablesValues || !Object.keys(agentInputVariablesValues).length) {
|
||||
agent = RunnableSequence.from([prompt, llm]).withConfig({
|
||||
metadata: { sequentialNodeName: agentName }
|
||||
})
|
||||
} else {
|
||||
agent = RunnableSequence.from([
|
||||
RunnablePassthrough.assign(transformObjectPropertyToFunction(agentInputVariablesValues, state)),
|
||||
prompt,
|
||||
llm
|
||||
]).withConfig({
|
||||
metadata: { sequentialNodeName: agentName }
|
||||
})
|
||||
}
|
||||
return agent
|
||||
} else {
|
||||
const promptArrays = [new MessagesPlaceholder('messages')] as BaseMessagePromptTemplateLike[]
|
||||
if (systemPrompt) promptArrays.unshift(['system', systemPrompt])
|
||||
if (humanPrompt) promptArrays.push(['human', humanPrompt])
|
||||
|
||||
const prompt = ChatPromptTemplate.fromMessages(promptArrays)
|
||||
|
||||
if (multiModalMessageContent.length) {
|
||||
const msg = HumanMessagePromptTemplate.fromTemplate([...multiModalMessageContent])
|
||||
prompt.promptMessages.splice(1, 0, msg)
|
||||
}
|
||||
|
||||
let conversationChain
|
||||
|
||||
if (!agentInputVariablesValues || !Object.keys(agentInputVariablesValues).length) {
|
||||
conversationChain = RunnableSequence.from([prompt, llm, new StringOutputParser()]).withConfig({
|
||||
metadata: { sequentialNodeName: agentName }
|
||||
})
|
||||
} else {
|
||||
conversationChain = RunnableSequence.from([
|
||||
RunnablePassthrough.assign(transformObjectPropertyToFunction(agentInputVariablesValues, state)),
|
||||
prompt,
|
||||
llm,
|
||||
new StringOutputParser()
|
||||
]).withConfig({
|
||||
metadata: { sequentialNodeName: agentName }
|
||||
})
|
||||
}
|
||||
|
||||
return conversationChain
|
||||
}
|
||||
}
|
||||
|
||||
async function agentNode(
|
||||
{
|
||||
state,
|
||||
llm,
|
||||
interrupt,
|
||||
agent,
|
||||
name,
|
||||
abortControllerSignal,
|
||||
nodeData,
|
||||
input,
|
||||
options
|
||||
}: {
|
||||
state: ISeqAgentsState
|
||||
llm: BaseChatModel
|
||||
interrupt: boolean
|
||||
agent: AgentExecutor | RunnableSequence
|
||||
name: string
|
||||
abortControllerSignal: AbortController
|
||||
nodeData: INodeData
|
||||
input: string
|
||||
options: ICommonObject
|
||||
},
|
||||
config: RunnableConfig
|
||||
) {
|
||||
try {
|
||||
if (abortControllerSignal.signal.aborted) {
|
||||
throw new Error('Aborted!')
|
||||
}
|
||||
|
||||
// @ts-ignore
|
||||
state.messages = restructureMessages(llm, state)
|
||||
|
||||
let result = await agent.invoke({ ...state, signal: abortControllerSignal.signal }, config)
|
||||
|
||||
if (interrupt) {
|
||||
const messages = state.messages as unknown as BaseMessage[]
|
||||
const lastMessage = messages[messages.length - 1]
|
||||
|
||||
// If the last message is a tool message and is an interrupted message, format output into standard agent output
|
||||
if (lastMessage._getType() === 'tool' && lastMessage.additional_kwargs?.nodeId === nodeData.id) {
|
||||
let formattedAgentResult: { output?: string; usedTools?: IUsedTool[]; sourceDocuments?: IDocument[] } = {}
|
||||
formattedAgentResult.output = result.content
|
||||
if (lastMessage.additional_kwargs?.usedTools) {
|
||||
formattedAgentResult.usedTools = lastMessage.additional_kwargs.usedTools as IUsedTool[]
|
||||
}
|
||||
if (lastMessage.additional_kwargs?.sourceDocuments) {
|
||||
formattedAgentResult.sourceDocuments = lastMessage.additional_kwargs.sourceDocuments as IDocument[]
|
||||
}
|
||||
result = formattedAgentResult
|
||||
} else {
|
||||
result.name = name
|
||||
result.additional_kwargs = { ...result.additional_kwargs, nodeId: nodeData.id, interrupt: true }
|
||||
return {
|
||||
messages: [result]
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
const additional_kwargs: ICommonObject = { nodeId: nodeData.id }
|
||||
|
||||
if (result.usedTools) {
|
||||
additional_kwargs.usedTools = result.usedTools
|
||||
}
|
||||
if (result.sourceDocuments) {
|
||||
additional_kwargs.sourceDocuments = result.sourceDocuments
|
||||
}
|
||||
if (result.output) {
|
||||
result.content = result.output
|
||||
delete result.output
|
||||
}
|
||||
|
||||
const outputContent = typeof result === 'string' ? result : result.content || result.output
|
||||
|
||||
if (nodeData.inputs?.updateStateMemoryUI || nodeData.inputs?.updateStateMemoryCode) {
|
||||
let formattedOutput = {
|
||||
...result,
|
||||
content: outputContent
|
||||
}
|
||||
const returnedOutput = await getReturnOutput(nodeData, input, options, formattedOutput, state)
|
||||
return {
|
||||
...returnedOutput,
|
||||
messages: convertCustomMessagesToBaseMessages([outputContent], name, additional_kwargs)
|
||||
}
|
||||
} else {
|
||||
return {
|
||||
messages: [
|
||||
new HumanMessage({
|
||||
content: outputContent,
|
||||
name,
|
||||
additional_kwargs: Object.keys(additional_kwargs).length ? additional_kwargs : undefined
|
||||
})
|
||||
]
|
||||
}
|
||||
}
|
||||
} catch (error) {
|
||||
throw new Error(error)
|
||||
}
|
||||
}
|
||||
|
||||
const getReturnOutput = async (nodeData: INodeData, input: string, options: ICommonObject, output: any, state: ISeqAgentsState) => {
|
||||
const appDataSource = options.appDataSource as DataSource
|
||||
const databaseEntities = options.databaseEntities as IDatabaseEntity
|
||||
const tabIdentifier = nodeData.inputs?.[`${TAB_IDENTIFIER}_${nodeData.id}`] as string
|
||||
const updateStateMemoryUI = nodeData.inputs?.updateStateMemoryUI as string
|
||||
const updateStateMemoryCode = nodeData.inputs?.updateStateMemoryCode as string
|
||||
|
||||
const selectedTab = tabIdentifier ? tabIdentifier.split(`_${nodeData.id}`)[0] : 'updateStateMemoryUI'
|
||||
const variables = await getVars(appDataSource, databaseEntities, nodeData)
|
||||
|
||||
const flow = {
|
||||
chatflowId: options.chatflowid,
|
||||
sessionId: options.sessionId,
|
||||
chatId: options.chatId,
|
||||
input,
|
||||
output,
|
||||
state,
|
||||
vars: prepareSandboxVars(variables)
|
||||
}
|
||||
|
||||
if (selectedTab === 'updateStateMemoryUI' && updateStateMemoryUI) {
|
||||
try {
|
||||
const parsedSchema = typeof updateStateMemoryUI === 'string' ? JSON.parse(updateStateMemoryUI) : updateStateMemoryUI
|
||||
const obj: ICommonObject = {}
|
||||
for (const sch of parsedSchema) {
|
||||
const key = sch.key
|
||||
if (!key) throw new Error(`Key is required`)
|
||||
let value = sch.value as string
|
||||
if (value.startsWith('$flow')) {
|
||||
value = customGet(flow, sch.value.replace('$flow.', ''))
|
||||
} else if (value.startsWith('$vars')) {
|
||||
value = customGet(flow, sch.value.replace('$', ''))
|
||||
}
|
||||
obj[key] = value
|
||||
}
|
||||
return obj
|
||||
} catch (e) {
|
||||
throw new Error(e)
|
||||
}
|
||||
} else if (selectedTab === 'updateStateMemoryCode' && updateStateMemoryCode) {
|
||||
const vm = await getVM(appDataSource, databaseEntities, nodeData, flow)
|
||||
try {
|
||||
const response = await vm.run(`module.exports = async function() {${updateStateMemoryCode}}()`, __dirname)
|
||||
if (typeof response !== 'object') throw new Error('Return output must be an object')
|
||||
return response
|
||||
} catch (e) {
|
||||
throw new Error(e)
|
||||
}
|
||||
}
|
||||
|
||||
return {}
|
||||
}
|
||||
|
||||
const convertCustomMessagesToBaseMessages = (messages: string[], name: string, additional_kwargs: ICommonObject) => {
|
||||
return messages.map((message) => {
|
||||
return new HumanMessage({
|
||||
content: message,
|
||||
name,
|
||||
additional_kwargs: Object.keys(additional_kwargs).length ? additional_kwargs : undefined
|
||||
})
|
||||
})
|
||||
}
|
||||
|
||||
class ToolNode<T extends BaseMessage[] | MessagesState> extends RunnableCallable<T, T> {
|
||||
tools: StructuredTool[]
|
||||
nodeData: INodeData
|
||||
inputQuery: string
|
||||
options: ICommonObject
|
||||
|
||||
constructor(
|
||||
tools: StructuredTool[],
|
||||
nodeData: INodeData,
|
||||
inputQuery: string,
|
||||
options: ICommonObject,
|
||||
name: string = 'tools',
|
||||
tags: string[] = [],
|
||||
metadata: ICommonObject = {}
|
||||
) {
|
||||
super({ name, metadata, tags, func: (input, config) => this.run(input, config) })
|
||||
this.tools = tools
|
||||
this.nodeData = nodeData
|
||||
this.inputQuery = inputQuery
|
||||
this.options = options
|
||||
}
|
||||
|
||||
private async run(input: BaseMessage[] | MessagesState, config: RunnableConfig): Promise<BaseMessage[] | MessagesState> {
|
||||
const message = Array.isArray(input) ? input[input.length - 1] : input.messages[input.messages.length - 1]
|
||||
|
||||
if (message._getType() !== 'ai') {
|
||||
throw new Error('ToolNode only accepts AIMessages as input.')
|
||||
}
|
||||
|
||||
const outputs = await Promise.all(
|
||||
(message as AIMessage).tool_calls?.map(async (call) => {
|
||||
const tool = this.tools.find((tool) => tool.name === call.name)
|
||||
if (tool === undefined) {
|
||||
throw new Error(`Tool ${call.name} not found.`)
|
||||
}
|
||||
let output = await tool.invoke(call.args, config)
|
||||
let sourceDocuments: Document[] = []
|
||||
if (output?.includes(SOURCE_DOCUMENTS_PREFIX)) {
|
||||
const outputArray = output.split(SOURCE_DOCUMENTS_PREFIX)
|
||||
output = outputArray[0]
|
||||
const docs = outputArray[1]
|
||||
try {
|
||||
sourceDocuments = JSON.parse(docs)
|
||||
} catch (e) {
|
||||
console.error('Error parsing source documents from tool')
|
||||
}
|
||||
}
|
||||
return new ToolMessage({
|
||||
name: tool.name,
|
||||
content: typeof output === 'string' ? output : JSON.stringify(output),
|
||||
tool_call_id: call.id!,
|
||||
additional_kwargs: {
|
||||
sourceDocuments,
|
||||
args: call.args,
|
||||
usedTools: [
|
||||
{
|
||||
tool: tool.name ?? '',
|
||||
toolInput: call.args,
|
||||
toolOutput: output
|
||||
}
|
||||
]
|
||||
}
|
||||
})
|
||||
}) ?? []
|
||||
)
|
||||
|
||||
const additional_kwargs: ICommonObject = { nodeId: this.nodeData.id }
|
||||
outputs.forEach((result) => (result.additional_kwargs = { ...result.additional_kwargs, ...additional_kwargs }))
|
||||
return Array.isArray(input) ? outputs : { messages: outputs }
|
||||
}
|
||||
}
|
||||
|
||||
module.exports = { nodeClass: Agent_SeqAgents }
|
||||
Reference in New Issue
Block a user