Voyage's contextualized embeddings accept a document's pre-split chunks and return one vector per chunk, so Qdrant can stay. Because vectors depend on neighboring text, I would re-embed the affected context group after edits, not cache by chunk text alone.

Context-aware vectors without replacing Qdrant

Checked September 29, 2026. Voyage's current documentation lists voyage-context-4 and accepts a nested list: one inner list per document, containing its pre-split chunks. Automatic vendor chunking is optional; keep it off to retain control of boundaries. Queries use the corresponding contextualized query encoding. Each chunk receives its own vector.

Jina's late-chunking method similarly aims to preserve surrounding context, by applying the transformer before pooling within chunk boundaries. These are related approaches, not an assertion of identical internals.

My proposed experiment: keep Git triggers, source reading, Qdrant and the backend unchanged, and compare independent chunk embeddings with contextualized ones. Respect the model's context limit; large documents may need bounded context groups.

The update implication is an inference: when a representation uses neighboring text, its cache key must cover that context and model configuration. Re-embedding changed files fits the reference. Reusing vectors for individually unchanged chunks requires additional dependency tracking. No retrieval improvement has been measured on this corpus.