Upgrade langchain-elasticsearch for LangChain 1.x compatibility

[sharrajesh]

Intent

Upgrade langchain-elasticsearch to support LangChain 1.x (released October 2025)
while maintaining backward compatibility with LangChain 0.3.x. This unblocks
downstream projects (like aiq_llm_apps) from upgrading to the latest LangChain
ecosystem.

---

Changes Made

1. pyproject.toml

| Change | From | To | Why
|
|----------------------|-----------------|----------------|------------------------
-------------------------------------------------|
| Version bump | 0.4.0 | 0.5.0 | Semantic versioning -
breaking change due to Python version requirement |
| Python version | >=3.9,<4.0 | >=3.10,<4.0 | LangChain 1.x requires
Python 3.10+ (Python 3.9 EOL Oct 2025) |
| langchain (test dep) | >=0.3.10,<1.0.0 | >=1.0.0,<2.0.0 | Enable testing against
LangChain 1.x |
| pytest-asyncio | ^0.21.1 | ^0.23.0 | Compatibility with
Python 3.10+ and newer pytest |

Note: langchain-core = ">=0.3.0,<2.0.0" was already compatible - no change needed.

---

2. Test Fixes (4 files)

Files:

• tests/unit_tests/_sync/test_cache.py
• tests/unit_tests/_async/test_cache.py
• tests/integration_tests/_sync/test_cache.py
• tests/integration_tests/_async/test_cache.py

Problem: Tests imported from langchain.embeddings.cache import _value_serializer -
this private module was reorganized in LangChain 1.x and no longer exists at that
path.

Solution: Replaced the import with a local implementation:
import json
from typing import List

def _value_serializer(value: List[float]) -> bytes:
"""Serialize embedding values to bytes (replaces private langchain
function)."""
return json.dumps(value).encode()

Why this approach:

• _value_serializer is a private function (underscore prefix) - not part of public API
• The implementation is trivial (json.dumps(value).encode())
• Avoids dependency on internal LangChain implementation details
• More robust against future LangChain changes

---

3. Import Path Fix (integration tests)

Change: from langchain.globals import set_llm_cache → from langchain_core.globals
import set_llm_cache

Why: langchain_core is the direct dependency of this package. Using
langchain.globals would require the full langchain package at runtime, but
langchain is only a test dependency. The langchain_core.globals path is the
canonical location.

---

4. Async Client Cleanup Fixture (async integration test)

Added: _close_async_caches autouse fixture in
tests/integration_tests/_async/test_cache.py

Why: Ensures async Elasticsearch clients are properly closed after tests, avoiding
aiohttp "Unclosed client session" warnings. This is a best practice for async tests
using aiohttp-based clients.

---

What Didn't Need Changes

The source code in langchain_elasticsearch/ was already compatible with LangChain
1.x because it uses stable langchain_core APIs:

• langchain_core.documents.Document
• langchain_core.embeddings.Embeddings
• langchain_core.vectorstores.VectorStore
• langchain_core.retrievers.BaseRetriever
• langchain_core.caches.BaseCache
• langchain_core.stores.ByteStore

---

Gotchas Found

1. Private API dependency in tests: The original tests depended on langchain.embeddings.cache._value_serializer, a private function. LangChain 1.x reorganized internal modules, breaking this import. Lesson: Avoid importing private APIs (underscore-prefixed) even in tests.
2. langchain vs langchain_core imports: Integration tests were importing from langchain.globals but langchain is only a test dependency. The correct import path is langchain_core.globals since that's the actual dependency.
3. Async client cleanup: Async Elasticsearch tests can leave unclosed aiohttp sessions if not properly cleaned up, causing warnings during test runs.
4. Python 3.9 EOL: LangChain 1.x dropped Python 3.9 support. This is a breaking change for any users still on Python 3.9.

---

Test Results

======================== 69 passed, 4 warnings in 0.13s ========================

The 4 warnings are expected deprecation notices for legacy retrieval strategy
classes (ApproxRetrievalStrategy → DenseVectorStrategy, etc.) which are still
functional.

---

Breaking Changes for Users

• Python 3.9 no longer supported - users must upgrade to Python 3.10+

---

[demetere]

Any timelines for this PR?

[rvargas42]

Please update this PR!