feat(embedding): surface non-symmetric embedding config for VikingDB provider

[mvanhorn]

Problem Statement

Non-symmetric embedding uses different representations for queries vs documents, improving retrieval quality for models that support it. The OpenAI, Gemini, Jina, and Minimax embedders already support query_param/document_param in ov.conf. The VikingDB embedder accepts is_query but ignores it -- all calls use symmetric mode regardless of config.

Closes #655.

Changes

• VikingDBDenseEmbedder: accept query_param/document_param, pass input_type to API data items
• VikingDBHybridEmbedder: same treatment
• VikingDBClientMixin._call_api(): accept optional input_type, add to request data items when set
• Factory entries: wire query_param/document_param from config to VikingDB embedder constructors
• Sparse embedder unchanged (sparse models are symmetric)

Config Example

[embedding.dense]
provider = "vikingdb"
model = "bge-m3"
query_param = "query"
document_param = "passage"

When configured, retrieval calls embed(text, is_query=True) which passes input_type=query in the API request. Indexing calls embed(text, is_query=False) which passes input_type=passage.

When not configured, behavior is unchanged (symmetric mode, no input_type in request).

Testing

4 unit tests:

• _resolve_input_type returns None in symmetric mode
• _resolve_input_type returns correct param for query vs document
• Hybrid embedder resolves correctly
• Backward compatibility without params

tests/unit/embedder/test_vikingdb_nonsymmetric.py ....  [100%]
4 passed

Implementation Notes

• Follows the existing pattern in openai_embedders.py:213-216
• _resolve_input_type() is a shared helper on Dense and Hybrid embedders
• The HierarchicalRetriever already passes is_query=True for queries (line 132), so this works end-to-end once configured

Feature Area

Retrieval/Search

This contribution was developed with AI assistance (Claude Code).

[CLAassistant]

All committers have signed the CLA.

[github-actions]

Failed to generate code suggestions for PR

[zhoujh01]

I think it might be more valuable to configure queries and documents as two separate models.

[mvanhorn]

The current implementation follows the query_param/document_param pattern from the OpenAI, Gemini, and Jina embedders (e.g., openai_embedders.py:213-216). That approach uses one model with different input type hints per call direction.

Are you thinking of a config where query and document each point to distinct VikingDB model names? Something like:

[embedding.dense]
provider = "vikingdb"
query_model = "bge-m3-query"
document_model = "bge-m3-passage"

Happy to restructure if that's the direction you prefer.

[MaojiaSheng]

need to fix

[mvanhorn]

Could you clarify what needs fixing? Happy to update the implementation if there are specific changes you'd like to see.

[MaojiaSheng]

Could you clarify what needs fixing? Happy to update the implementation if there are specific changes you'd like to see.

sorry, I mean the conflicts. Thanks