STORM Retrieval
KEY POINT
Self-Improving RAG
Automates fine-tuning using feedback, enabling continuous improvement in retrieval quality through simple user response evaluations.
Optimized Retrieval Architecture for Large Document Sets
Multi-stage filtering efficiently navigates large document collections, delivering precise answers even within complex datasets.
Improved Accuracy through Hybrid Search
Combines lexical search for exact keyword matching with semantic search for contextual similarity, ensuring accurate and comprehensive results.
RAG PERFORMANCE
By optimizing each engine in the tech stack and designing a high-efficiency vector search logic, we have significantly enhanced the performance and quality of our RAG system

1. Compared the search quality of <O Corp. Embedding + Re-ranker> and <STORM Embedding + Re-ranker + Search Logic> using RECALL@K.
2. Compared the search quality of <O Corp. Embedding> and <STORM Embedding> using RECALL@K.
3. After retrieving chunks using the <O Corp. embedding model>, compared the re-ranking performance between <O Corp. Re-ranker> and <STORM Re-ranker> using RECALL@K.
CORE TECHNOLOGY

Optimized Chunking Strategy
Enhances contextual understanding by preserving document-level flow and leveraging multi-scale chunking.
- LATE CHUNKING: Maintains global document context
- MULTI-SCALE CHUNKING: Utilizes chunks of multiple sizes simultaneously
- CONTEXTUAL CHUNKING: Adds local context to each chunk

Query Expansion
nterprets the intent behind user questions and enriches the query with synonyms and semantically related expressions to improve retrieval accuracy and coverage.
- Analyzes diverse cues in the query and explores multiple retrieval paths
- Automatically expands to include semantically related terms and synonyms
- Reformulates queries based on prior conversation context
- Interprets and adjusts for temporal expressions like “last year” or “yesterday”

Hybrid Search
Maximizes search accuracy through LEXICAL + SEMANTIC retrieval strategies.
- Exact keyword matching (LEXICAL)
- Semantic similarity search (SEMANTIC)
- Combined and optimized hybrid result ranking

Post-Processing
Optimizes final results through multi-reranking and chunk clustering.
- Re-ranks outputs based on semantic similarity
- Improving LLM response quality by grouping related chunks to enhance context
- Merges overlapping and adjacent information to generate more refined inputs, enabling high-quality LLM responses