Migrating from 0.3 to 0.4

The 0.4 line preserves the public 0.3 call shapes. A composite client with chat() and embed() is still accepted everywhere, and data written without a scope remains available under its original doc_id.

Separate chat and embeddings

Existing code does not need to change:

pipeline = Pipeline(store, llm)

When providers differ, pass the two capabilities explicitly:

pipeline = Pipeline(
    store,
    chat_client=chat_client,
    embedding_client=embedding_client,
)

LLMClientProtocol remains the composite of ChatClientProtocol and EmbeddingClientProtocol. CompositeLLMClient can assemble one explicitly.

Enabling MemoryScope

Physical keys do not change when scope is omitted. For an incremental migration, create scoped components only for new tenants or threads:

scope = MemoryScope(tenant="acme", user="u-42", thread="chat-7")
indexer = DocumentIndexer(store, embedding_client, scope=scope)
builder = ContextBuilder(store, embedding_client, scope=scope)

Records created without a scope are not copied into the new namespace automatically. To migrate existing memory, read it through an unscoped object and re-index it through a scoped indexer. Do not pair a scoped writer with an unscoped reader: no scope deliberately means the legacy/global namespace.

For direct StoreProtocol deletion, use scoped_doc_id(logical_id, scope). High-level components map IDs automatically.

Checking a custom adapter

The contract kit does not depend on pytest:

from protoprompt.testing import check_embedding_client, check_vector_store

await check_embedding_client(my_embedding_client)
await check_vector_store(my_isolated_test_store)

Checks use temporary keys and clean them up. A dedicated test collection or schema is still recommended for a production backend.