fix(vector-io): replace inline::milvus with inline::sqlite-vec as default (#301)

[RHAIENG-3916](https://redhat.atlassian.net/browse/RHAIENG-3916)

# What does this PR do?

Milvus-lite acquires an exclusive file lock on startup, which prevents any second process from opening the database. This causes the LlamaStack pod to crash when configured with multiple workers. This PR replaces `inline::milvus` with `inline::sqlite-vec` as the always-enabled default vector_io provider.

The `remote::milvus` provider is unaffected.

## Test Plan

Manual verification only. Automated multi-worker vector_io testing is not included in this PR because `inline::sqlite-vec` currently opens SQLite without WAL mode or a busy timeout, which makes a 2-worker test unreliable under concurrent writes. [5344](ogx-ai/ogx#5344 (comment)) has been filed to address [`_create_sqlite_connection`](https://github.com/llamastack/llama-stack/blob/f7a10a4748a19979214f5a9d642945cc9b2ce36f/src/llama_stack/providers/inline/vector_io/sqlite_vec/sqlite_vec.py#L109); a smoke test covering multi-worker vector_io will be added once that fix lands.

- [X] Build the container image and verify `sqlite-vec` installs cleanly
- [X] Start the server and confirm vector_io operations work correctly with a single worker
- [X] Verify `remote::milvus` activates when `MILVUS_ENDPOINT` is set



## Summary by CodeRabbit

* **Changes**
  * Switched the default vector database backend from Milvus to SQLite-Vec.
  * Updated container dependencies and vector storage configuration to support the new backend.

* **Documentation**
  * Updated provider documentation to reflect the new default vector database.




Approved-by: rhdedgar

Approved-by: cdoern

Author: mergify[bot]

distribution/Containerfile

@@ -20,7 +20,6 @@ RUN uv pip install \
     'mcp>=1.23.0' \
     'nltk>=3.9.4' \
     'pymilvus==2.6.9' \
-    'pymilvus[milvus-lite]==2.6.9' \
     'pypdf>=6.7.2' \
     aiosqlite \
     asyncpg \
@@ -49,6 +48,7 @@ RUN uv pip install \
     scipy \
     sentencepiece \
     sqlalchemy[asyncio] \
+    sqlite-vec \
     tokenizers \
     tqdm \
     transformers \

distribution/README.md

@@ -38,7 +38,7 @@ You can see an overview of the APIs and Providers the image ships with in the ta
 | tool_runtime | remote::model-context-protocol | No | ✅ | N/A |
 | tool_runtime | remote::tavily-search | No | ✅ | N/A |
 | vector_io | inline::faiss | No | ❌ | Set the `ENABLE_FAISS` environment variable |
-| vector_io | inline::milvus | No | ✅ | N/A |
+| vector_io | inline::sqlite-vec | No | ✅ | N/A |
 | vector_io | remote::milvus | No | ❌ | Set the `MILVUS_ENDPOINT` environment variable |
 | vector_io | remote::pgvector | No | ❌ | Set the `ENABLE_PGVECTOR` environment variable |
 | vector_io | remote::qdrant | No | ❌ | Set the `ENABLE_QDRANT` environment variable |

distribution/config.yaml

@@ -58,13 +58,13 @@ providers:
       api_key: ${env.OPENAI_API_KEY:=}
       base_url: ${env.OPENAI_BASE_URL:=https://api.openai.com/v1}
   vector_io:
-  - provider_id: milvus
-    provider_type: inline::milvus
+  - provider_id: sqlite-vec
+    provider_type: inline::sqlite-vec
     config:
-      db_path: /opt/app-root/src/.llama/distributions/rh/milvus.db
+      db_path: /opt/app-root/src/.llama/distributions/rh/sqlite_vec.db
       persistence:
         backend: kv_default
-        namespace: vector_io::milvus
+        namespace: vector_io::sqlite_vec
   - provider_id: ${env.ENABLE_FAISS:+faiss}
     provider_type: inline::faiss
     config: