---
id: "concept-model-driven-retrieval"
type: "concept"
source_timestamps: ["00:07:45", "00:08:35"]
tags: ["rag", "data-architecture", "system-design"]
related: ["prereq-rag-architecture", "question-model-driven-tool-architecture"]
definition: "An evolution of RAG where hardcoded semantic search logic is abandoned, allowing the LLM itself to navigate, query, and filter raw data repositories."
sources: ["s44-claude-mythos"]
sourceVaultSlug: "s44-claude-mythos"
originDay: 44
---
# Model-Driven Retrieval

## Definition

An evolution of [[prereq-rag-architecture|Retrieval-Augmented Generation]] in which **hardcoded semantic search logic is abandoned**, and the LLM itself navigates, queries, and filters raw data repositories.

## What it replaces

Traditional RAG hardcoded:
- Semantic search algorithms
- Chunking strategies
- Ranking mechanisms
- Pre-determined context selection

Humans engineered all of these to feed "the right" context to the model.

## The shift

With massive context windows and superior reasoning (the capability claim attached to [[concept-claude-mythos]]), the model can do this work itself, more accurately than human-engineered pipelines. The architecture becomes:

1. Expose a well-organized, searchable repository — file system, codebase, raw database.
2. Let the model decide:
   - What to query
   - How to search
   - What to pull into working memory
3. Trust the model to recognize its own knowledge gaps.

## Where the industry is heading

Related industry threads:
- Toolformer (Schick et al., 2023) — LLMs invoking APIs autonomously
- Gorilla (Xia et al., 2023) — LLM-driven tool use over hardcoded RAG
- Devin / Cognition Labs — file-system-native agents

## Open architectural question

How do we expose a multi-terabyte enterprise database to an LLM efficiently without overwhelming the context window or causing hallucinated queries? See [[question-model-driven-tool-architecture]].

## Position in the framework

Step 3 of the [[framework-mythos-readiness|Mythos Readiness Transformation]] — *"Architect for Tools."* Pairs with [[concept-outcome-driven-prompting]] (model decides the *how*) and [[concept-single-eval-gate]] (no intermediate checks on retrieval choices).