What

Adds an optional ingest_pipeline parameter to ElasticsearchDocumentStore. When set, write_documents and write_documents_async pass it as the Elasticsearch bulk API pipeline argument so ingest pipelines (including inference processors) can run at index time. Default behavior is unchanged (None).

Why

• Fixes Index-time inference in ElasticsearchDocumentStore (optional / deferred) #2940

Users who index into Elasticsearch with a pre-defined ingest pipeline had no way to attach that pipeline from Haystack. The parent discussion in #699 defers broader retrieval work; this change is the narrow index-time hook only.

How

• New ingest_pipeline on ElasticsearchDocumentStore.__init__: non-empty strings only (after strip); whitespace-only raises ValueError.
• helpers.bulk / helpers.async_bulk receive pipeline=... only when the value is set, so deletes and existing callers are unaffected.
• to_dict / from_dict include the field like sparse_vector_field (missing key deserializes to None).

Testing

Unit tests cover serialization (test_to_dict, from_dict variants), validation, and mocked bulk calls to assert pipeline is passed or omitted. Retriever to_dict / from_dict expectations were updated for nested document store serialization.

To reproduce locally:

cd integrations/elasticsearch
hatch run test:unit
hatch run test:types

For formatting on Windows, if hatch run fmt fails on path handling, use:

hatch -e default run ruff check --fix <paths>
hatch -e default run ruff format <paths>

Test output:

117 passed, 182 deselected (unit tests on current main)

Type check:

Success: no issues found in 10 source files

Trade-offs

• No per-call pipeline override on write_documents (only store-level). Can be added later if needed.
• Pipeline creation and mapping design stay in Elasticsearch; Haystack does not validate that the pipeline exists.