Store embedding locally

Problem/Motivation

Currently, the module triggers a re-computation of vector embeddings whenever the index configuration is saved or updated, even if the changes do not affect the content being embedded. This behavior is inefficient and costly because:

• Generating embeddings (e.g., via external APIs like OpenAI) often incurs financial costs per token.
• The process is computationally expensive and slows down the indexing operation.
• It is unnecessary to regenerate embeddings if the source fields (e.g., title, body) and the embedding configuration (e.g., model selection) have not changed.

Steps to reproduce

1. Configure a Search API index using the Typesense backend with embeddings enabled on specific fields (e.g., body).
2. Index the content to generate the initial embeddings.
3. Go to the index schema configuration page and change a setting unrelated to embeddings (e.g., change the "Batch size" or disable a different, unrelated processor).
4. Save the configuration.
5. Observe that the system marks the items for re-indexing and subsequently re-computes the embeddings, leading to unnecessary API calls.

Proposed resolution

Modify the logic to detect whether the configuration changes actually impact the embedding generation. The system should only trigger a re-computation of embeddings if:

• The fields selected for embedding have been modified.
• The embedding configuration itself (e.g., the chosen model or strategy) has changed.

If the configuration change is irrelevant to the embeddings (e.g., changing a different processor), the existing embeddings should be preserved to save resources.

Remaining tasks

• Review the code changes in the merge request.
• Test to ensure that changing unrelated index configuration does not trigger embedding re-computation.
• Test to verify that changing the embedding settings does trigger a re-computation as expected.

User interface changes

None.

API changes

None.

Data model changes

None.