---
id: "question-model-driven-tool-architecture"
type: "open-question"
source_timestamps: ["00:08:15"]
tags: ["system-architecture", "tooling"]
related: ["concept-model-driven-retrieval"]
resolution_path: "The release of new frameworks and best practices for exposing file systems and databases directly to LLMs without intermediate semantic search layers."
sources: ["s44-claude-mythos"]
sourceVaultSlug: "s44-claude-mythos"
originDay: 44
---
# Architectures for Model-Driven Retrieval at Scale

## The question

**How do we expose massive, multi-terabyte enterprise data to an LLM in a way that supports [[concept-model-driven-retrieval|Model-Driven Retrieval]] at scale — without overwhelming context windows or causing hallucinated queries?**

Sub-questions:
- What tool-use interface design lets an LLM navigate a database without seeing the full schema in-context?
- How do we balance model autonomy in retrieval against query cost and latency?
- How do we audit / explain retrieval decisions when the model is making them?
- What replaces semantic-search infrastructure when retrieval becomes model-driven?

## Why it matters

The speaker advocates abandoning hardcoded RAG, but provides no detailed architectural blueprint. The industry needs new standards for tool-use interfaces to support this paradigm at production scale.

## Resolution path

Watch for:
- New frameworks for direct file-system / database exposure to LLMs (post-Toolformer, post-Gorilla)
- Best-practice patterns from labs deploying agents on large enterprise datasets
- MCP (Model Context Protocol) and similar standardization efforts
- Benchmarks comparing model-driven vs hardcoded retrieval at multi-TB scale

## Related work mentioned in enrichment

- Toolformer (Schick et al., 2023, arXiv:2302.04761)
- Gorilla (Xia et al., 2023)
- Devin (Cognition Labs) — file-system-native agent

## Related

- Concept: [[concept-model-driven-retrieval]]
- Prerequisite: [[prereq-rag-architecture]]