langchain[minor]: Add pgvector to list of supported vectorstores in s…

…elf query retriever (#22678)

The fact that we outsourced pgvector to another project has an
unintended effect. The mapping dictionary found by
`_get_builtin_translator()` cannot recognize the new version of pgvector
because it comes from another package.
`SelfQueryRetriever` no longer knows `PGVector`.

I propose to fix this by creating a global dictionary that can be
populated by various database implementations. Thus, importing
`langchain_postgres` will allow the registration of the `PGvector`
mapping.

But for the moment I'm just adding a lazy import

Furthermore, the implementation of _get_builtin_translator()
reconstructs the BUILTIN_TRANSLATORS variable with each invocation,
which is not very efficient. A global map would be an optimization.

- **Twitter handle:** pprados

@eyurtsev, can you review this PR? And unlock the PR [Add async mode for
pgvector](langchain-ai/langchain-postgres#32)
and PR [community[minor]: Add SQL storage
implementation](#22207)?

Are you in favour of a global dictionary-based implementation of
Translator?

libs/langchain/langchain/retrievers/self_query/base.py

@@ -168,6 +168,14 @@ def _get_builtin_translator(vectorstore: VectorStore) -> Visitor:
             if isinstance(vectorstore, Chroma):
                 return ChromaTranslator()
 
+        try:
+            from langchain_postgres import PGVector
+        except ImportError:
+            pass
+        else:
+            if isinstance(vectorstore, PGVector):
+                return PGVectorTranslator()
+
         raise ValueError(
             f"Self query retriever with Vector Store type {vectorstore.__class__}"
             f" not supported."