Query Augmentation is the First Layer of Agentic RAG
victorialslocum · x · 2026-07-15
The core message here is: no matter how good your retrieval is, it cannot save a misunderstood user query. If the input itself is ambiguous, lacks context, or contains multiple intents, subsequent steps like retrieval, memory, tools, and agents will only amplify the error.
The author considers query augmentation as the first step in agentic RAG, outlining three traditional methods:
- Rewriting: Rephrasing the query for clarity to improve retrieval matching.
- Expansion: Adding related terms and synonyms to cover intents not explicitly stated by the user.
- Decomposition: Breaking down complex questions into multiple sub-questions before aggregating the answers.
She further emphasizes that a query agent is the truly effective approach: letting an LLM dynamically decide how to rewrite, expand, or decompose, rather than relying on hardcoded rules. The post notes that Weaviate's Query Agent reads collection schemas, selects relevant collections, constructs queries with filters and aggregations, and checks if the results actually answer the question—rewriting further if not—while maintaining conversational context during follow-ups.
The conclusion is clear: in context engineering, query augmentation is the absolute frontline layer; all subsequent retrieval, memory, tool use, and agent reasoning depend on it getting the question right first.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21