← Home
TWIML · June 9, 2026 · 51m

Is RAG Dead? Lessons from Building AI for Tax Law with Alex Bowcut

Alex Bowcut
Show Notes

Alex Bowcut, Head of Engineering at Sphere, challenges the narrative that retrieval-augmented generation (RAG) is obsolete now that context windows accommodate millions of tokens. He explains why RAG remains critical in high-stakes domains like global tax compliance, where accuracy alone isn't enough—every AI conclusion must be backed by precise legal citations and withstand expert scrutiny. Sphere's TRAM system combines retrieval, reasoning models, legal review workflows, and reinforcement learning to help tax experts work nearly two orders of magnitude faster while maintaining trustworthiness and accuracy.

This summary was generated from show notes and public descriptions, not from a full transcript review. Details may contain inaccuracies.

Novel

•
TRAM: Hybrid System Combining Retrieval, Reasoning, Expert Feedback, and Reinforcement Learning
Sphere's TRAM system integrates retrieval, reasoning models, legal review workflows, reinforcement learning, and deterministic logic to achieve near-100x productivity gains for tax experts while maintaining accuracy and auditability.
ClaudeChatGPTGemini

Highlights

•
RAG Remains Critical in High-Stakes Domains
Despite massive context windows, retrieval-augmented generation is essential in regulated domains like tax law because accuracy requires precise legal citations and expert-verifiable sources, not just correct answers.
ClaudeChatGPTGemini
•
Retrieval Architecture Tradeoffs: Semantic Chunking, Dense vs. Sparse Retrieval
Processing legal documents requires careful choices about chunking strategies, semantic boundaries, and hybrid retrieval methods—dense embeddings for meaning, sparse methods for precision—with no universal winner.
ClaudeChatGPTGemini

Editorial

•
Citation Accuracy as Non-Negotiable Requirement
In regulated domains like tax law, being able to point to the exact legal source of a decision is as important as having the right answer—sometimes more important.
ClaudeChatGPTGemini
•
Multi-Jurisdictional Complexity Requires Deterministic Logic and Auditability
Tax regulations vary across jurisdictions with conflicting rules and local exceptions, forcing Sphere to build systems that enforce precedence hierarchies and can explain jurisdiction-specific logic—a problem pure LLMs struggle to solve reliably.
ClaudeChatGPTGemini

Misc

✧Tax compliance is a domain where AI mistakes have immediate, measurable legal and financial consequences
✧The phrase 'two orders of magnitude faster' suggests a 100x productivity gain for tax experts in a regulated domain
✧Sphere processes tax regulations across multiple jurisdictions globally
✧Context window size is not the primary constraint; retrieval precision, citation accuracy, and auditability are
Was this useful?