Semantic Query Parsing
Semantic query interpretation for LLM-based retrieval and document search systems.
Overview
This project implements a semantic query parsing pipeline to convert raw user queries into structured representations for downstream tasks such as Agentic RAG (see project).
Implementation Details
- Linguistic Preprocessing: tokenization, normalization, part-of-speech tagging
- Intent Modeling: transformer-based multi-intent classification, hierarchical labels
- Entity Extraction: NER, domain-specific linking, slot filling
- Semantic Representation: JSON, logical forms, graph structures
- Synthetic Query Generation: LLM-guided paraphrasing, slot-preserving data augmentation
- Similarity & Retrieval: Bi-encoder embeddings for retrieval, cross-encoder re-ranking
Integration with Agentic RAG
- Parsed semantic queries are fed into the Planner Agent for task decomposition
- Enables structured retrieval, grounded reasoning, and multi-step planning
- Works alongside Prompt Engineering & Evaluation for instruction optimization and observability
Use Cases
- Knowledge-intensive search engines
- Conversational agents with precise intent understanding
- Recommendation systems
- Analytics for query behavior and clustering