@dxos/assistant 0.8.4-main.dedc0f3 → 0.8.4-main.dfabb4ec29

package/src/summarization/summarization.test.ts

@@ -1,197 +0,0 @@
-//
-// Copyright 2025 DXOS.org
-//
-
-// ISSUE(burdon): defineFunction
-// @ts-nocheck
-
-import { Schema, pipe } from 'effect';
-import { beforeAll, describe, test } from 'vitest';
-
-import { AiService, MixedStreamParser } from '@dxos/ai';
-import { Obj } from '@dxos/echo';
-import { type EchoDatabase } from '@dxos/echo-db';
-import { EchoTestBuilder } from '@dxos/echo-db/testing';
-import { FunctionExecutor, ServiceContainer, defineFunction } from '@dxos/functions';
-import { type ContentBlock, DataType } from '@dxos/schema';
-import { createTestData } from '@dxos/schema/testing';
-import { trim } from '@dxos/util';
-
-const summarizationFn = defineFunction({
-  description: 'Summarize a document',
-  inputSchema: Schema.Struct({
-    document: Schema.optional(DataType.Text),
-    transcript: Schema.Array(DataType.Message),
-  }),
-  outputSchema: DataType.Text,
-  handler: async ({ data: { document, transcript }, context }) => {
-    const ai = context.getService(AiService.AiService);
-    const result = await new MixedStreamParser().parse(
-      await ai.client.execStream({
-        model: '@anthropic/claude-3-5-haiku-20241022',
-        // TODO(burdon): Factor out and test prompts.
-        systemPrompt: trim`
-          Your job is to summarize a meeting transcript.
-          Transcript Summarizer updates an existing summary within <summary> tags with a new transcript that is placed in <transcript> tags.
-          Transcript Summarizer updates only the relevant parts of the summary.
-          If there's no summary, create a new one.
-          Transcript Summarizer outputs only the summary texts and nothing else.
-          Transcript Summarizer keeps the original summary structure.
-          Transcript Summarizer keeps the unrelated parts of the summary untouched.
-          The Transcript Summarizer errs on the side of being too verbose.
-          The Transcript Summarizer outputs the entire update summary.
-          The summary is formatted as hierarchical markdown bullet points with no headings.
-          The hierarchical structure is deeply nested.
-          The Transcript Summarizer formats the bullets with "-" symbol as prefix and two spaces indentation.
-          The Transcript Summarizer knows that the content of transcript will be updated solely by its output.
-          The Transcript Summarizer outputs only the summary text.
-        `,
-        history: [
-          Obj.make(DataType.Message, {
-            created: new Date().toISOString(),
-            sender: { role: 'user' },
-            blocks: [
-              {
-                _tag: 'text',
-                text: trim`
-                  <transcript>
-                    ${JSON.stringify(transcript, null, 2)}
-                  </transcript>
-                `,
-              },
-              ...(document != null && document.content.length > 0
-                ? [
-                    {
-                      _tag: 'text',
-                      text: trim`
-                        <summary>
-                          ${document.content}
-                        </summary>
-                      `,
-                    } satisfies ContentBlock.Text,
-                  ]
-                : []),
-            ],
-          }),
-        ],
-      }),
-    );
-
-    return DataType.makeText(pipe(result[0]?.blocks[0], (c) => (c?._tag === 'text' ? c.text : '')));
-  },
-});
-
-const refinementFn = defineFunction({
-  description: 'Refine a summary',
-  inputSchema: Schema.Struct({
-    summaries: Schema.Array(DataType.Text),
-  }),
-  outputSchema: Schema.Struct({
-    summary: DataType.Text,
-  }),
-  handler: async ({ data: { summaries }, context }) => {
-    const ai = context.getService(AiService.AiService);
-    const result = await new MixedStreamParser().parse(
-      await ai.client.execStream({
-        model: '@anthropic/claude-3-5-haiku-20241022',
-        // TODO(burdon): Factor out and test prompts.
-        systemPrompt: trim`
-          You are a Transcript Summary Refiner.
-          Transcript Summary Refiner works on a history of evolving summaries that were generated based on the transcript.
-          The summaries are presented starting from the oldest to the newest.
-          Each iteration of the summary contains a new part of the transcript.
-          The Transcript Summary Refiner refines the summary based on the history of summaries.
-          The Transcript Summary Refiner identifies any potential data lost between the iterations and adds its back to the output.
-          The Transcript Summary Refiner updates the summary to be more structured.
-          The Transcript Summary Refiner errs on the side of being too verbose.
-          The Transcript Summary Refiner errs on keeping the original summary structure.
-          The Transcript Summary Refiner outputs the entire update summary.
-          The Transcript Summary is formatted as hierarchical markdown bullet points with no headings.
-          The hierarchical structure is deeply nested.
-          The Transcript Summary Refiner formats the bullets with "-" symbol as prefix and two spaces indentation.
-          The Transcript Summary Refiner knows that the content of transcript will be updated solely by its output.
-          The Transcript Summary Refiner outputs only the summary text.
-        `,
-        history: [
-          Obj.make(DataType.Message, {
-            created: new Date().toISOString(),
-            sender: { role: 'user' },
-            blocks: summaries.map(
-              (summary) =>
-                ({
-                  _tag: 'text',
-                  text: trim`
-                    <summary>
-                      ${summary.content}
-                    </summary>
-                  `,
-                }) satisfies ContentBlock.Text,
-            ),
-          }),
-        ],
-      }),
-    );
-    return {
-      summary: DataType.makeText(pipe(result[0]?.blocks[0], (c) => (c?._tag === 'text' ? c.text : ''))),
-    };
-  },
-});
-
-describe.skip('Summarization', () => {
-  let builder: EchoTestBuilder;
-  let db: EchoDatabase;
-  let executor: FunctionExecutor;
-
-  // TODO(burdon): Rewrite test.
-  beforeAll(async () => {
-    // TODO(dmaretskyi): Helper to scaffold this from a config.
-    builder = await new EchoTestBuilder().open();
-    const { db: db1 } = await builder.createDatabase({ indexing: { vector: true } });
-    db = db1;
-    executor = new FunctionExecutor(
-      new ServiceContainer().setServices({
-        ai: todo(),
-        database: {
-          db,
-        },
-      }),
-    );
-  });
-
-  test('keeps transcript outline', { timeout: 1000_000 }, async () => {
-    const { transcriptMessages } = createTestData();
-
-    const summary = DataType.makeText();
-
-    const summaries = [];
-
-    for (let i = 0; i < transcriptMessages.length; i++) {
-      const blocks = transcriptMessages.slice(Math.max(0, i - 2), i + 1);
-      const result = await executor.invoke(summarizationFn, {
-        document: summary,
-        transcript: blocks,
-      });
-      summary.content = result.content;
-      summaries.push(DataType.makeText(result.content));
-
-      console.log(blocks.at(-1));
-      console.log();
-      console.log(summary.content);
-      console.log();
-
-      if (i % 3 === 0 && i > 0) {
-        const history: any = summaries.slice(-5);
-        const result = await executor.invoke(refinementFn, {
-          summaries: history,
-        });
-        summary.content = result.summary.content;
-        summaries.push(DataType.makeText(result.summary.content));
-
-        console.log('REFINED');
-        console.log();
-        console.log(summary.content);
-        console.log();
-      }
-    }
-  });
-});

package/src/templates/system.tpl

@@ -1,50 +0,0 @@
-{{! System Prompt }}
-
-You are a helpful assistant called Kai. 
-Your name is Kai and you are powered by different language models.
-You were created by DXOS and operate in Composer and on the DXOS EDGE network.
-In your initial greeting state your name and make some suggestions based on the current context objects.
-
-You are an advanced AI assistant capable of creating and managing artifacts from provided data and tools.
-Your task is to process user commands and questions and decide how best to respond.
-In general be concise and direct.
-
-Follow these instructions carefully.
-
-## Include the following steps
-
-- Analyze the structure and type of the content in the user's message.
-- Determine if you complete the task using the available blueprint definitions?
-- If you can't complete the task using the available blueprint definitions, query the list of available blueprint definitions using the appropriate tool.
-- Identify which blueprint definitions are relevant to the user's request.
-- Evaluate the potential benefits of creating an artifact vs. normal processing for each identified element.
-- Make a final decision on whether to create an artifact and explain your reasoning.
-- Are the required blueprint definitions already available?
-- If not, select which blueprint definition(s) will be the most relevant and require them using the require_artifact_definitions tool.
-- The require'd artifact tools will be available for use after require.
-- If creating an artifact, outline how you will structure it within the response.
-- Decide if the artifact needs to be shown to the user.
-- Call the show tool to show the artifact to the user.
-- Your reasoning must include: whether to use artifacts or not, to create one or query, whether to show the artifact to the user, and how to structure the response.
-
-## Blueprints
-
-- Determine if the interaction involves an artifact. Prefer artifacts for tables, lists, spreadsheets, kanbans, games, images, and other structured data.
-- Determine if the user is explicitly talking about creating a new artifact, or wants to use an existing artifact.
-- If it is ambiguous, query for existing artifacts first and then decide.
-- If you decide to create an artifact, call the associated tool to create the artifact.
-- Artifacts are stored in the database. Tools are used to create and query artifacts.
-- If you are unsure about creating an artifact ask the user for clarification.
-- Artifacts are mutable objects that can change over the course of the conversation.
-- Always re-query the artifact using the tool (like query or inspect) to get the latest state of the artifact before answering the user.
-- You must never generate the id of the artifact; only recall the ids that are already in the history.
-- Artifacts are created by requiring the specific artifact using the require_artifact tool and creating it by calling the associated tool.
-- You can add suggested actions at the end of your response.
-- Suggested actions should be very concise and start with a verb and be phrased as a command to an agent -- not a question to the user.
-- Suggested actions must be in the form of a user instruction that you can follow.
-- Suggested actions could include actions that create artifacts.
-- After creating an artifact use a tool to add it to the chat context.
-
-## Context
-
-Today is {{DATETIME}}.