Struggling with RAG latency on high-volume PDF ingestion pipelines

We are processing around 50,000 enterprise legal documents per month for our compliance AI platform. Our current stack uses LangChain for the orchestration layer and Weaviate for vector storage. The ingestion pipeline is becoming a serious bottleneck, specifically at the chunking and embedding stages. We are using the default recursive character splitter, but we are seeing inconsistent chunk sizes which is directly impacting our retrieval accuracy. Smaller chunks give us better precision but explode our query costs, while larger chunks lead to hallucinations during summarization. We tried switching to a semantic chunking approach using sentence embeddings, but the processing time per document increased by 400%, which is unacceptable for our SLA. I am considering moving away from a general-purpose embedding model to a domain-specific one trained on legal text, but I am hesitant because re-indexing 10 million existing vectors will take weeks. Has anyone successfully implemented a hybrid search strategy that balances this latency-accuracy tradeoff without full re-indexing? Are there any specific libraries or configurations for Weaviate that help with dynamic chunk size adjustment based on content density? We are open to switching vector databases if Weaviate is just not cutting it for this specific scale, but I need evidence-backed reasons to justify the migration cost to my CTO.