fix(milvus): make get_milvus_client directory-create loop-safe for milvus-lite

[sahilkanger]

fix(milvus): make get_milvus_client directory-create loop-safe for milvus-lite

Refs #113

Summary

MilvusAdapter.get_milvus_client() crashes when used with a local (milvus-lite) file URI
from inside cognee's running event loop. It tried to ensure the parent directory exists by
driving cognee's async file-storage helper from a sync method: loop.run_until_complete(...)
raises RuntimeError on the already-running loop, and the except RuntimeError fallback calls
asyncio.run(...), which also raises inside a running loop. This PR replaces that block with a
plain, loop-safe os.makedirs(db_dir, exist_ok=True).

Root cause

get_milvus_client() is synchronous but is called from async adapter methods
(create_collection, etc.) that run inside cognee's event loop. Both branches of the old
directory-ensure logic touch the event loop:

loop = asyncio.get_event_loop()
loop.run_until_complete(file_storage.ensure_directory_exists())   # RuntimeError: loop already running
except RuntimeError:
    asyncio.run(file_storage.ensure_directory_exists())           # RuntimeError: asyncio.run() from running loop

So any local milvus-lite run fails before a collection can be created. Server URIs (http...)
skip the block and were unaffected.

Solution

A local milvus-lite URI is always a real filesystem path, so ensuring its parent directory needs
no async storage abstraction:

if not self.url.startswith("http"):
    db_dir = os.path.dirname(self.url)
    if db_dir:
        os.makedirs(db_dir, exist_ok=True)

os was already imported; the now-unused get_file_storage import is removed. Version bumped
0.1.2 → 0.1.3 per the repo's fix-bump convention.

Scope note (the other half of #113)

The issue also surfaces a KeyError: 'milvus' under ENABLE_BACKEND_ACCESS_CONTROL=True, because
Milvus ships no DatasetDatabaseHandler. I deliberately left that out of this PR: while
investigating I found the adapter doesn't currently scope its connection to database_name (the
MilvusClient is created without db_name) and prune() drops every collection on the
connection — so simply mirroring the Qdrant handler (dataset → vector_database_name, delete →
prune()) would make deleting one dataset wipe all of them. Doing multi-tenant Milvus correctly
needs a real per-dataset isolation model (server db_name create/drop vs. per-file milvus-lite)
that should be validated against a live instance. I've left a separate comment on #113 to align
with maintainers on the intended isolation model before building it, rather than ship a data-loss
risk here. Happy to follow up with that PR once the design is agreed.

Changes

0cd6d0a fix(milvus): make get_milvus_client directory-create loop-safe for milvus-lite

 .../milvus_adapter.py                 | 22 +++++++---------------
 packages/vector/milvus/pyproject.toml |  2 +-
 2 files changed, 8 insertions(+), 16 deletions(-)

Testing performed

• python -m py_compile on the changed module — OK.
• ruff check / ruff format --check on the package — clean.
• Not run: the package's CI scripts (examples/example.py, tests/test_milvus.py) exercise this
  path with a local URI, but need LLM/embedding keys and (for the example) a reachable Milvus;
  those secrets aren't available to a fork PR locally. The change is a self-contained,
  loop-safe directory create.

Checklist

• Change is focused and minimal
• Follows the project's code style / conventions
• Tests added or updated where appropriate (CI validates via example/test scripts; see Testing)
• Documentation updated where appropriate (n/a — no public API change)
• lint passes locally (ruff check on the package)
• format passes locally (ruff format --check on the package)

[sahilkanger]

Two notes on CI: (1) Milvus Adapter Test fails with LLMAPIKeyNotSetError — it runs the example/test scripts which need the repo's LLM secret, unavailable to fork PRs, so it fails before reaching get_milvus_client. (2) the ruff Run Formatting Check red is the repo-wide ruff-latest drift (0.16.0 reformats ~25 pre-existing files) — I opened #147 to pin ruff and fix that for all PRs. This change itself is lint/format-clean. Glad to approve-and-run if that helps validate.