fix(core): preserve precomputed embeddings in async multimodal text retrieval

[tteok1]

Description

Preserve a supplied QueryBundle.embedding in MultiModalVectorIndexRetriever.atext_retrieve, matching synchronous text retrieval. Previously, the async path replaced it with a newly computed vector, changing the retrieved document, or raised ValueError when only an embedding was supplied.

New Package?

• Yes
• No

Version Bump?

Not applicable to llama-index-core.

Type of Change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

• Added unit tests covering supplied embeddings with and without query text, normal text embedding, custom embedding strings, and an empty text store.
• Before the fix: 2 regression cases failed and 3 control cases passed. After the fix: all 5 passed.
• Related retriever, vector-store and multimodal chat tests: 45 passed, including the 5 new cases.
• Core Mypy: no issues in 479 source files.
• Ruff lint/format checks and git diff --check: passed.
• Repository pre-commit checks passed on both changed files.
• The standalone reproduction now returns the same document for sync and async retrieval.

Local validation used Windows/Python 3.12.

AI assistance

AI assisted with the guard change and regression tests. Validation used a real in-memory index/vector store with fixed embedding models, with the checks listed above run locally.