[upstream PR 225] feat(embedding): configurable local model, cache dir, and HF mirror | 嵌入模型可配置

[wbugitlab1]

Source: Source pull request number: 225 in rohitg00/agentmemory (URL omitted to avoid GitHub cross-reference)
Title: feat(embedding): configurable local model, cache dir, and HF mirror | 嵌入模型可配置
Author: mechanic-Q
State: open
Draft: no
Merged: no
Head: mechanic-Q/agentmemory:feature/configurable-embedding @ a8f9d89
Base: main @ 1c8713f
Labels: (none)
Changed files: 0
Commits: 0
Created: 2026-05-02T07:26:35Z
Updated: 2026-05-17T09:42:42Z
Closed: (not closed)
Merged at: (not merged)

Original PR body:

Summary | 概述

Make the local embedding provider configurable via environment variables, enabling multilingual embedding models (bge-m3, multilingual-e5, etc.) and supporting users behind firewalls or in regions with restricted HuggingFace access.

通过环境变量使本地嵌入模型可配置，支持切换多语言模型和配置 HuggingFace 镜像。

Motivation | 动机

The local embedding provider was hardcoded to all-MiniLM-L6-v2 (384-dim English-only). Users with non-English content (Chinese, Japanese, Korean, multilingual codebases) had no way to use a multilingual embedding model without forking the codebase. Additionally, users behind corporate firewalls or in regions with slow/unreachable HuggingFace access had no way to configure a mirror endpoint.

本地嵌入模型被硬编码为 all-MiniLM-L6-v2（仅支持英文）。非英文用户无法切换多语言模型。部署在防火墙后或在 HuggingFace 受限地区的用户无法配置镜像。

Changes | 改动

Three new environment variables (all optional, defaults unchanged):

• AGENTMEMORY_LOCAL_EMBEDDING_MODEL — model name (default: Xenova/all-MiniLM-L6-v2). Set to Xenova/bge-m3 for multilingual 1024-dim embeddings supporting 100+ languages including Chinese.
• AGENTMEMORY_LOCAL_EMBEDDING_MODEL_DIR — custom cache directory for pre-downloaded models (useful in air-gapped/offline environments).
• HF_ENDPOINT — HuggingFace mirror endpoint (e.g. https://hf-mirror.com).

Auto-detects dimensions from a known list; falls back to 384 for unknown models.

Usage | 用法

# Multilingual Chinese support via bge-m3
AGENTMEMORY_LOCAL_EMBEDDING_MODEL=Xenova/bge-m3

# Offline/pre-downloaded models
AGENTMEMORY_LOCAL_EMBEDDING_MODEL_DIR=/opt/models/transformers

# HF mirror for China/firewall users
HF_ENDPOINT=https://hf-mirror.com

Backwards Compatibility | 向后兼容

All env vars are optional. Default behavior unchanged — all-MiniLM-L6-v2 with default HF endpoint.

Summary by CodeRabbit

• New Features
  • Embedding provider now supports dynamic model and dimension configuration through environment variables instead of fixed defaults
  • Users can specify a custom local model cache directory for embedding models
  • Added support for Hugging Face endpoint configuration

Local branch:
Fork PR:
Fork decision:
Verification:
Notes: