Convex Agents Context

Sstobo

更新日 Today

83 閲覧

メタai

について

このスキルは、開発者がLLMが各生成で受け取るコンテキストをカスタマイズできるようにし、メッセージ履歴やRAGコンテキスト注入を制御できます。クロススレッド検索、メモリ注入、会話要約、メッセージフィルタリングなどの高度なパターンをサポートしています。トークンオーバーフローを防止したり、ユーザープロファイルを追加したり、少数ショットの例でLLMの動作を誘導するために使用できます。

クイックインストール

Claude Code

推奨

プラグインコマンド推奨

/plugin add https://github.com/Sstobo/convex-skills

Git クローン代替

git clone https://github.com/Sstobo/convex-skills.git ~/.claude/skills/Convex Agents Context

このコマンドをClaude Codeにコピー＆ペーストしてスキルをインストールします

ドキュメント

Purpose

By default, the Agent includes recent messages as context. This skill covers customizing that behavior for advanced patterns like cross-thread search, memory injection, summarization, and filtering.

When to Use This Skill

Limiting context window to prevent token overflow
Searching across multiple threads for relevant context
Injecting memories or user profiles into every prompt
Summarizing long conversations before continuing
Filtering out sensitive or irrelevant messages
Adding few-shot examples to guide LLM

Configure Default Context Options

const myAgent = new Agent(components.agent, {
  name: "My Agent",
  languageModel: openai.chat("gpt-4o-mini"),
  contextOptions: {
    recentMessages: 50,
    excludeToolMessages: true,
    searchOptions: {
      limit: 10,
      textSearch: true,
      vectorSearch: false,
    },
  },
});

Override Context Per Call

export const generateWithCustomContext = action({
  args: { threadId: v.string(), prompt: v.string() },
  handler: async (ctx, { threadId, prompt }) => {
    const result = await myAgent.generateText(
      ctx,
      { threadId },
      { prompt },
      {
        contextOptions: {
          recentMessages: 20,
          searchOptions: {
            limit: 5,
            textSearch: true,
            vectorSearch: true,
          },
        },
      }
    );

    return result.text;
  },
});

Search Across Threads

export const generateWithCrossThreadContext = action({
  args: { threadId: v.string(), userId: v.string(), prompt: v.string() },
  handler: async (ctx, { threadId, userId, prompt }) => {
    const result = await myAgent.generateText(
      ctx,
      { threadId, userId },
      { prompt },
      {
        contextOptions: {
          searchOtherThreads: true,
          searchOptions: {
            limit: 15,
            textSearch: true,
            vectorSearch: true,
          },
        },
      }
    );

    return result.text;
  },
});

Custom Context Handler

Completely customize context:

const myAgent = new Agent(components.agent, {
  name: "My Agent",
  languageModel: openai.chat("gpt-4o-mini"),
  contextHandler: async (ctx, args) => {
    const userMemories = await getUserMemories(ctx, args.userId);
    const examples = getExamples();

    return [
      ...userMemories,
      ...examples,
      ...args.search,
      ...args.recent,
      ...args.inputMessages,
    ];
  },
});

Fetch Context Manually

Get context without calling LLM:

import { fetchContextWithPrompt } from "@convex-dev/agent";

export const getContextForPrompt = action({
  args: { threadId: v.string(), prompt: v.string() },
  handler: async (ctx, { threadId, prompt }) => {
    const { messages } = await fetchContextWithPrompt(ctx, components.agent, {
      threadId,
      prompt,
      contextOptions: {
        recentMessages: 20,
        searchOptions: { limit: 10, textSearch: true },
      },
    });

    return messages;
  },
});

Key Principles

Default context is sufficient: Most use cases work with defaults
Search improves relevance: Enable for long conversations
userId required for cross-thread: Provide when searching multiple threads
Context handlers are powerful: Use for memories, examples, special formatting
Recent messages take precedence: Used after search in context order

Next Steps

See rag for knowledge base context injection
See fundamentals for agent setup
See rate-limiting for token management

GitHub リポジトリ

Sstobo/convex-skills

パス: convex-agents-context