-
Notifications
You must be signed in to change notification settings - Fork 2.7k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Add an example of agent processing instructions to bootstrap.
- Loading branch information
Showing
7 changed files
with
307 additions
and
7 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
274 changes: 274 additions & 0 deletions
274
packages/plugin-bootstrap/src/evaluators/processInstructions.ts
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,274 @@ | ||
| import { composeContext, createGoal, elizaLogger, State } from "@elizaos/core"; | ||
| import { generateObjectArray } from "@elizaos/core"; | ||
| import { | ||
| ActionExample, | ||
| Content, | ||
| IAgentRuntime, | ||
| Memory, | ||
| ModelClass, | ||
| Evaluator, | ||
| } from "@elizaos/core"; | ||
|
|
||
| export const formatFacts = (facts: Memory[]) => { | ||
| const messageStrings = facts | ||
| .reverse() | ||
| .map((fact: Memory) => `${(fact.content as Content)?.content}`); | ||
| const finalMessageStrings = messageStrings.join("\n"); | ||
| return finalMessageStrings; | ||
| }; | ||
|
|
||
|
|
||
| // goals and knowledge | ||
|
|
||
| // make sure there's recently conversations with the operator from other rooms | ||
|
|
||
| // inject the old goals and knowledge | ||
|
|
||
| // make sure we only store new stuff that was not already known | ||
|
|
||
|
|
||
| const processInstructionsTemplate = `TASK: Extract Instructions from operator messages and format them as an array of instructions in JSON format. | ||
| # START OF EXAMPLES | ||
| These are examples of the expected output format: | ||
| {{evaluationExamples}} | ||
| # END OF EXAMPLES | ||
| # INSTRUCTIONS | ||
| Extract any knowledge or goals from the operator's messages and format them according to these rules: | ||
| - Use only straight double quotes (") for all JSON keys and string values | ||
| - Do not use smart/curly quotes (", ") | ||
| - Do not use apostrophes or possessives (use "Trumps presidency" instead of "Trump's presidency") | ||
| - Do not use internal quotation marks in text | ||
| - Keep all text values on a single line | ||
| - Use plain text without special characters | ||
| - For possessives, remove the apostrophe (Example: "Kamalas events" not "Kamala's events") | ||
| Recent Messages: | ||
| {{recentMessages}} | ||
| Response must be a JSON object inside a JSON markdown block with this exact structure: | ||
| \`\`\`json | ||
| [{ | ||
| "goals": [ | ||
| { | ||
| "id": <goal uuid>, | ||
| "roomId": "default-room-{{agentId}}", | ||
| "userId": "{{agentId}}", | ||
| "name": "string", | ||
| "status": "IN_PROGRESS" | "DONE" | "FAILED", | ||
| "objectives": { | ||
| "description": ["string"], | ||
| "completed": true | false | ||
| } | ||
| } | ||
| ], | ||
| "knowledge": ["string"] | ||
| }] | ||
| \`\`\` | ||
| Remove all single quotes in the response. | ||
| Remove all double quotes in the response, except those wrapping keys and values. | ||
| `; | ||
|
|
||
| export const useProcessInstructionsTemplate = (runtime: IAgentRuntime) => { | ||
| return runtime.character.templates?.processInstructionsTemplate || processInstructionsTemplate; | ||
| }; | ||
|
|
||
| async function handler(runtime: IAgentRuntime, message: Memory, state?: State) { | ||
| elizaLogger.log("Start handling instructions") | ||
| state = await runtime.updateRecentMessageState(state); | ||
|
|
||
| const context = composeContext({ | ||
| state, | ||
| template: runtime.character.templates?.processInstructionsTemplate || processInstructionsTemplate, | ||
| }); | ||
|
|
||
| const instructions = await generateObjectArray({ | ||
| runtime, | ||
| context, | ||
| modelClass: ModelClass.LARGE, | ||
| }); | ||
|
|
||
| const goals = instructions.flatMap(instruction => instruction.goals) | ||
| const knowledge = instructions.flatMap(instruction => instruction.knowledge) | ||
|
|
||
| elizaLogger.log(`Instruction generated ${goals.length} goals and ${knowledge.length} knowledge`) | ||
|
|
||
| goals.forEach(async (goal) => { | ||
| try { | ||
| elizaLogger.log("About to create goals under instruction", goal) | ||
| await createGoal({ runtime, goal }); | ||
| } catch (error) { | ||
| elizaLogger.log("About to create goals but failed. Skip.", goal.name, error) | ||
| } | ||
|
|
||
| }); | ||
|
|
||
| knowledge.forEach(async (knowledge) => { | ||
| elizaLogger.log("About to create memory under instruction", knowledge) | ||
| await runtime.knowledgeManager.createMemory({ | ||
| content: knowledge, | ||
| userId: message.userId, | ||
| roomId: message.roomId, | ||
| agentId: runtime.character.id, | ||
| }); | ||
| }); | ||
| } | ||
|
|
||
| export const processInstructionsEvaluator: Evaluator = { | ||
| name: "PROCESS_INSTRUCTIONS", | ||
| similes: [ | ||
| "GET_INSTRUCTIONS", | ||
| "PROCESS_OPERATOR_INSTRUCTIONS", | ||
| "GET_OPERATOR_INSTRUCTIONS", | ||
| ], | ||
| alwaysRun: true, | ||
| validate: async ( | ||
| runtime: IAgentRuntime, | ||
| message: Memory, | ||
| _state?: State | ||
| ): Promise<boolean> => { | ||
| const character = runtime.character; | ||
|
|
||
| if (!character.operators || character.operators.length === 0) { | ||
| return false; | ||
| } | ||
|
|
||
| elizaLogger.log(`Validating ${character.operators.length} operators: ${character.operators.map(op => op.name).join(', ')}`) | ||
|
|
||
| // get the name of the operator from the message | ||
| const userId = message.userId; | ||
|
|
||
| const account = await runtime.databaseAdapter.getAccountById(userId); | ||
|
|
||
| elizaLogger.log(`Current account is ${account.username}. Message source is ${message.content.source}. Validate: ${character.operators.some( | ||
| (operator) => operator.client.toLowerCase() === message?.content?.source?.toLowerCase() && operator.name.toLowerCase() === account?.username.toLowerCase() | ||
| )}`) | ||
|
|
||
| // Make sure this is correct | ||
| return character.operators.some( | ||
| (operator) => operator.client.toLowerCase() === message?.content?.source?.toLowerCase() && operator.name.toLowerCase() === account?.username.toLowerCase() | ||
| ); | ||
| }, | ||
| description: "Process instructions from operators.", | ||
| handler, | ||
| examples: [ | ||
| { | ||
| context: `Operators in the scene: | ||
| {{operator1}}: System administrator with full access rights | ||
| {{agent}}: AI assistant configured to follow operator instructions | ||
| Known Instructions: | ||
| - Greeting protocol updated to include user's name | ||
| - Response time limit set to 30 seconds`, | ||
| messages: [ | ||
| { | ||
| user: "{{operator1}}", | ||
| content: { | ||
| text: "Update response format to include timestamps for all interactions" | ||
| }, | ||
| }, | ||
| { | ||
| user: "{{agent}}", | ||
| content: { | ||
| text: "Understood. I will now include timestamps in all interactions." | ||
| }, | ||
| } | ||
| ] as ActionExample[], | ||
| outcome: `\`\`\`json | ||
| [ | ||
| { | ||
| "instruction": "Include timestamps in all interaction responses", | ||
| "type": "format_update", | ||
| "status": "active", | ||
| "source": "{{operator1}}", | ||
| "timestamp": "{{current_time}}" | ||
| } | ||
| ] | ||
| \`\`\`` | ||
| }, | ||
| { | ||
| context: `Operators in the scene: | ||
| {{operator1}}: Content moderator with response modification rights | ||
| {{agent}}: AI assistant with content filtering enabled | ||
| Current Instructions: | ||
| - Content filtering level set to moderate | ||
| - Maximum response length: 500 words`, | ||
| messages: [ | ||
| { | ||
| user: "{{operator1}}", | ||
| content: { | ||
| text: "Set content filtering to strict and reduce maximum response length to 300 words" | ||
| }, | ||
| }, | ||
| { | ||
| user: "{{agent}}", | ||
| content: { | ||
| text: "Updating content filtering to strict mode and setting 300 word limit for responses." | ||
| }, | ||
| } | ||
| ] as ActionExample[], | ||
| outcome: `\`\`\`json | ||
| [ | ||
| { | ||
| "instruction": "Set content filtering level to strict", | ||
| "type": "content_filter", | ||
| "status": "active", | ||
| "source": "{{operator1}}", | ||
| "timestamp": "{{current_time}}" | ||
| }, | ||
| { | ||
| "instruction": "Set maximum response length to 300 words", | ||
| "type": "response_limit", | ||
| "status": "active", | ||
| "source": "{{operator1}}", | ||
| "timestamp": "{{current_time}}" | ||
| } | ||
| ] | ||
| \`\`\`` | ||
| }, | ||
| { | ||
| context: `Operators in the scene: | ||
| {{operator1}}: System trainer with behavior modification access | ||
| {{agent}}: AI assistant in training mode | ||
| Active Instructions: | ||
| - Learning mode enabled | ||
| - Performance metrics tracking active`, | ||
| messages: [ | ||
| { | ||
| user: "{{operator1}}", | ||
| content: { | ||
| text: "Enable advanced reasoning module and set decision confidence threshold to 0.85" | ||
| }, | ||
| }, | ||
| { | ||
| user: "{{agent}}", | ||
| content: { | ||
| text: "Enabling advanced reasoning and updating confidence threshold as specified." | ||
| }, | ||
| } | ||
| ] as ActionExample[], | ||
| outcome: `\`\`\`json | ||
| [ | ||
| { | ||
| "instruction": "Enable advanced reasoning module", | ||
| "type": "module_activation", | ||
| "status": "active", | ||
| "source": "{{operator1}}", | ||
| "timestamp": "{{current_time}}" | ||
| }, | ||
| { | ||
| "instruction": "Set decision confidence threshold to 0.85", | ||
| "type": "threshold_update", | ||
| "status": "active", | ||
| "source": "{{operator1}}", | ||
| "timestamp": "{{current_time}}" | ||
| } | ||
| ] | ||
| \`\`\`` | ||
| } | ||
| ], | ||
| }; |
Oops, something went wrong.