[hulincup]

Linked issue: #958

Purpose of change

The YAML loader resolves short clazz: aliases through three hand-maintained copies of the same alias table (Java Aliases.java, Python aliases.py, docs yaml.md). Six resource implementations already have a ResourceName constant but no alias entry in any of the three tables, so users must reach them by fully-qualified class path. Per item 1 of #958, this adds the missing entries across all three tables, mirroring the shape of existing entries (e.g. the Bedrock aliases added in #812).

Alias	Resource type	Language	Implementation
gemini	chat model connection + setup	java	Gemini
azure_openai	chat model connection + setup	java	Azure OpenAI
opensearch	vector store	java	OpenSearch
s3_vectors	vector store	java	S3 Vectors
milvus	vector store	java	Milvus
mem0	vector store	python	Mem0

All additions are purely additive. Java/Python/API-YAML alignment is preserved (the same six entries land in the Java table, the Python table, and the docs table).

Scope: item 1 of #958 only. Item 2 — the azure alias resolving to the Azure AI Inference connection (AzureAIChatModelConnection) while yaml.md documents it as "Azure OpenAI (Java)" — is a behavior change for users who already have clazz: azure in their YAML and is intentionally left for a separate PR, as the issue suggests. The existing azure alias and its docs row are untouched here. Adding azure_openai as a new Java alias is independent of that decision and gives the Azure OpenAI Java connection a reachable short name either way.

Open question from the issue: should any of these six be deliberately unaliased? I defaulted to aliasing all six, since an alias is an optional short name (fully-qualified clazz: still works) and adding one is reversible. Happy to drop any a maintainer feels should stay first-class-only.

Tests

• python/flink_agents/api/yaml/tests/test_aliases.py: added two cases asserting the six new aliases resolve to the expected FQNs. For azure_openai — the only alias whose Java/Python simple names collide (AzureOpenAIChatModelConnection) — the assertions pin the package prefix (startswith("org.apache.flink.agents")) in addition to the simple name, so a Java entry that lost its .Java suffix and resolved to the Python class would fail. Full file passes locally: 13 passed.
• api/src/test/java/org/apache/flink/agents/api/yaml/AliasesTest.java: added two parallel cases mirroring the existing clazzAliasJavaBucket style.
• ruff check clean on the changed Python file.

Java build was not run locally (sparse checkout of a large repo; host JDK is 8 vs the project Java 11 target). The change is mechanical — each new entry mirrors the immediately adjacent existing put/dict lines — and the referenced ResourceName constants all exist at main. CI will confirm.

API

No public API signature changes. Aliases are optional short names; fully-qualified clazz: values keep passing through unchanged.

Documentation

• doc-included

docs/content/docs/development/yaml.md is updated in this PR: the azure_openai row gains a type: java column entry, and mem0 / opensearch / s3_vectors / milvus rows are added to the vector-store alias table.

Part of #958 (item 1).

Generated-by: Claude Code (Anthropic) — used to author this patch and PR description; reviewed and submitted by a human contributor.

[weiqingy]

Thanks for taking this on. Only one inline comment below.

FYI in case it's useful, since you deliberately scoped item 2 out: it has a doc-only PR up in #963, so that half is covered and this one gets to stay purely additive.

[weiqingy]

AzureOpenAIChatModelConnection is the simple name on both sides of the bridge:

• Python: ...chat_models.azure.azure_openai_chat_model.AzureOpenAIChatModelConnection
• Java: ...integrations.chatmodels.openai.AzureOpenAIChatModelConnection

So this passes even if the java entry ever lost its .Java and resolved to the Python class, and the same goes for the Setup assertion two lines down. The Java mirror catches that case because isEqualTo compares the whole constant, so the two sides aren't guarding equally right now.

test_resolve_clazz_dispatches_on_language above already pins the package for this reason. Would it be worth borrowing that here?

azure_conn = resolve_clazz("azure_openai", ResourceType.CHAT_MODEL_CONNECTION, "java")
assert azure_conn.startswith("org.apache.flink.agents")
assert azure_conn.endswith("AzureOpenAIChatModelConnection")

azure_openai is the only one of the six where the names collide, so it would just be these two.

[hulincup]

Good catch, thanks. ab251a7 pins the package on both the connection and setup assertions for azure_openai — startswith("org.apache.flink.agents") plus the existing endswith, mirroring test_resolve_clazz_dispatches_on_language. It's the only one of the six where the Java/Python simple names collide, so the other five keep the endswith-only form. 13 passed locally on test_aliases.py.

[hulincup]

Good catch — done in 78b8328. azure_openai now asserts both startswith("org.apache.flink.agents") and endswith(...) on the connection and the setup, mirroring test_resolve_clazz_dispatches_on_language. 13 passed locally, including test_resolve_clazz_covers_chat_model_java_gemini_and_azure_openai.

[weiqingy]

One note on the body: GitHub's closing-keyword parser stops at the number, so Closes #958 (item 1). doesn't actually scope to item 1. Merging would close all of #958, taking item 2 and the still-open question about whether any of the six should stay deliberately unaliased along with it. Would Part of #958 (item 1). do the job here? Same link, minus the auto-close.

One other thing worth folding in: the repo keeps a PR template at .github/PULL_REQUEST_TEMPLATE.md asking for a Linked issue: line plus ### Purpose of change, ### Tests, ### API, and ### Documentation. Easy to miss, since GitHub only pre-fills it when you open the PR through the web form. Your body already covers most of it, so it's mostly remapping headings, and doc-included looks like the right box given the yaml.md change.

[hulincup]

Thanks @weiqingy, both points addressed in ab251a7:

1. Closing keyword — switched Closes #958 (item 1). to Part of #958 (item 1). so merging doesn't auto-close [Bug] Six resource implementations have no YAML alias, and the azure alias documents the wrong provider #958 and drag item 2 (and the open "should any stay unaliased" question) along with it.

2. PR template — remapped the body to the .github/PULL_REQUEST_TEMPLATE.md headings (Linked issue, ### Purpose of change / ### Tests / ### API / ### Documentation) and ticked doc-included, since yaml.md is updated in-tree.

[wenjin272]

Thanks for taking this on @hulincup. LGTM.

Sorry, I merged #963 first without noticing the overlap, so this PR now needs a conflict resolution.

[hulincup]

Thanks @wenjin272 and @weiqingy for the review.

Conflict resolved — rebased onto main to absorb #963. The yaml.md overlap is gone: kept #963's Azure AI (Java) wording on the azure row and layered the new gemini row + the azure_openai Java column on top. Force-pushed; CI is rerunning (mergeable is back to MERGEABLE).

@weiqingy's top-level points — both addressed:

• Closing keyword switched to Part of #958 (item 1). so merging won't auto-close [Bug] Six resource implementations have no YAML alias, and the azure alias documents the wrong provider #958 (item 2 + the open question stay visible).
• Body remapped to the PR template sections (Linked issue, ### Purpose of change / Tests / API / Documentation) with doc-included checked.

@weiqingy's line comment on test_aliases.py — addressed in 78b8328; replying in that thread.