[Bug] L0 embedding silently fails — MAX_BATCH_SIZE=256 exceeds Tencent API limit of 10

[yrwd999]

OpenClaw Version | OpenClaw 版本

2026.6.8

Plugin Version | 插件版本

0.3.6

Operating System | 操作系统

Linux

System Specification | 系统配置

No response

Describe the bug | 问题描述

L0 capture path 的背景 embedding 任务持续失败，但仅记录为 non-fatal WARN，导致向量 embedding 数据永久丢失，混合搜索降级为纯关键词搜索。

根因：src/core/store/embedding.ts 第 374–375 行硬编码 MAX_BATCH_SIZE = 256，而dashscope embedding API 严格限制每批最多 10 条文本。当 L0 捕获的文本数在 11–256 之间时，embedBatch() 内部的 chunking 条件 processedTexts.length > MAX_BATCH_SIZE 不触发，批次直接以 >10 条发送至 API，返回 HTTP 400。

To Reproduce | 复现步骤

1. 安装 memory-tencentdb 插件，配置 Dashscope embedding
2. 正常使用 Agent 一段时间
3. 观察日志：[memory-tdai] [capture] [L0-vec-index-bg] Background embedding failed (non-fatal)
4. 查询数据库差值持续扩大

Expected behavior | 预期行为

• 方案 A：MAX_BATCH_SIZE 应 ≤ 10，确保始终触发 chunking
• 方案 B：batch size > 10 时应拆分为多个 ≤10 的子批次
• 方案 C：API 返回 400 后应重试（指数退避）而非直接丢弃

Error Logs / Screenshots | 报错日志/截图

[memory-tdai] [capture] [L0-vec-index-bg] Background embedding failed (614ms, non-fatal): Embedding API error: HTTP 400 Bad Request —
{
"error": {
"message": "<400> InternalError.Algo.InvalidParameter: Value error, batch size is invalid, it should not be larger than 10.: input.contents",
"type": "InvalidParameter",
"code": "InvalidParameter"
}
}

Additional context | 补充信息

影响范围：

• L0 向量索引写入 0% 成功率（持续至今，约 2026-05-26 起）
• L1/L2/L3 管线未受影响
• recall 降级：hybrid 搜索退化为 keyword-only，语义相似度匹配完全失效

代码位置：src/core/store/embedding.ts:374-483

• 第 374 行：const MAX_BATCH_SIZE = 256;
• 第 480–483 行：仅在 processedTexts.length > MAX_BATCH_SIZE 时才进行 chunking
• 当文本数 11 ≤ n ≤ 256：chunking 逻辑不触发

建议修复：

1. 高优：将 MAX_BATCH_SIZE 改为 10
2. 次优：在 auto-capture.ts 中先手动 chunk 为 ≤10 的子批次

[Maxwell-Code07]

Thanks for the report! Clear root cause — MAX_BATCH_SIZE=256 exceeds dashscope's 10-per-batch limit. YOMXXX has already submitted PR #238 with the fix, we'll review it shortly

[YOMXXX]

Triage update:

Kept #238 as the focused verified fix for the remote embedding batch-size rejection: cap remote batches at 10 and preserve existing chunking behavior.

#238 verification:

• COREPACK_ENABLE_AUTO_PIN=0 pnpm vitest run src/core/store/embedding.test.ts
• COREPACK_ENABLE_AUTO_PIN=0 pnpm test
• COREPACK_ENABLE_AUTO_PIN=0 pnpm build
• git diff --check

#424 has a potentially useful configurable embedding.batchSize extension, but it should be split/rebased as a follow-up on top of #238 instead of merging as a competing broad fix for the same immediate bug.