feat: add Valkey as RAG vector store backend

[daric93]

Closes #2062

Summary

Adds Valkey (via the Valkey Search module) as a RAG vector store backend in MetaGPT, following the existing ConfigBasedFactory pattern used by FAISS, Chroma, and Elasticsearch.

Changes

• Schema (metagpt/rag/schema.py): ValkeyStoreConfig, ValkeyIndexConfig, ValkeyRetrieverConfig
• Vector store (metagpt/rag/vector_stores/valkey.py): ValkeyVectorStore implementing llama-index BasePydanticVectorStore, backed by valkey-glide FT.CREATE / FT.SEARCH (HNSW + KNN)
• Retriever (metagpt/rag/retrievers/valkey_retriever.py): ValkeyRetriever extending RAGRetriever
• Factories: registered Valkey configs in RAGIndexFactory and RetrieverFactory (lazy imports so valkey-glide stays optional)
• Dependencies (setup.py): valkey-glide>=2.1.0,<3.0.0 added to the rag extra
• Config example (config/config2.example.yaml): commented Valkey section

Design notes

• Storage: JSON documents with an HNSW VectorField, COSINE distance, FLOAT32. FLAT algorithm also supported.
• client_name="metagpt_rag_client" on all connections.
• Sync-over-async handled by a single persistent event loop on a dedicated thread, so the cached client always runs on its creation loop.
• Batch insert via asyncio.gather (_BATCH_SIZE=100); raises on partial failure.
• SCAN loops bounded by _MAX_SCAN_ITERATIONS; drop_index() always cleans orphaned keys.

Testing

• 54 tests passing (unit + integration) on Python 3.11 in a venv.
• 14 live integration tests run against a live Valkey container (valkey/valkey-bundle:9.1, Search module) on localhost:6379 — not skipped.
  • 11 vector store integration tests + 3 retriever integration tests
• 40 unit tests (mocked valkey-glide client) covering connect/TLS/password/timeout, index creation, batch insert + partial-failure, KNN query, delete, drop-index cleanup, SCAN safety limit, and persistent-loop reuse.
• black, isort, ruff all clean.
• Existing FAISS/Chroma/ES/BM25 factory tests still pass — purely additive, no regressions.

How to test manually

# Start Valkey with the Search module
podman run -d --name valkey-test -p 6379:6379 valkey/valkey-bundle:9.1

# Run the suite
ALLOW_OPENAI_API_CALL=0 python -m pytest \
  tests/metagpt/rag/vector_stores/ \
  tests/metagpt/rag/retrievers/test_valkey_retriever.py \
  tests/metagpt/rag/retrievers/test_valkey_retriever_integration.py \
  tests/metagpt/rag/factories/test_index.py \
  tests/metagpt/rag/factories/test_retriever.py \
  -v --timeout=60

Configuration required

New optional valkey section in config2.yaml (see config/config2.example.yaml):

Key	Default	Description
host	localhost	Valkey server host
port	6379	Valkey server port
password	None	Optional auth password
use_tls	false	TLS for cloud deployments (e.g. ElastiCache)
request_timeout	5000	Request timeout (ms)
index_name	metagpt_rag	FT.SEARCH index name
prefix	metagpt:rag:	Key prefix for document storage
vector_dimensions	1536	Embedding dimensionality
distance_metric	COSINE	COSINE, L2, or IP
vector_algorithm	HNSW	HNSW or FLAT

Requirements

Requires a Valkey deployment with the Search module (e.g. valkey/valkey-bundle:9.1). Purely additive — no changes to existing backends.