The platform team will fund exactly one index for a 500...

Do not arbitrate based on team preference; instead, focus on the arithmetic of the storage requirements and the failure modes of the models. At 500 million chunks, a dense index using 768-dimensional fp32 embeddings would require approximately 1.54 TB of storage. Furthermore, dense retrieval is fundamentally ill-suited for the half of your traffic consisting of part numbers and error codes. These are rare tokens that are typically poorly represented in embedding spaces, leading to poor performance for exact-match retrieval requirements.