Query Expansion

Leverages embedding models to find semantically related terms or phrases for expansion. The query is embedded, and its nearest neighbors in the embedding space are identified and added.

• Process: Generate embedding for the query → perform k-Nearest Neighbors (k-NN) search in a vocabulary embedding space → add neighboring terms.
• Advantage: Captures semantic relatedness beyond strict synonymy (e.g., "ML" → "machine learning," "neural networks," "training data").
• Implementation: Can use general models (Sentence-BERT) or domain-fine-tuned models. Often integrated directly with vector database query logic.

Uses a large language model (LLM) to rewrite or expand the original query. The LLM is prompted to generate multiple paraphrases, related questions, or keyword expansions based on the query's intent and assumed context.

• Techniques:
  • Paraphrasing: "How to optimize a SQL query?" → "What are techniques for improving SQL performance?"
  • Hypothetical Document Embedding (HyDE): The LLM generates a hypothetical ideal answer; this document is then embedded and used as the query for vector search.
  • Step-Back Prompting: LLM generates a more general, conceptual query to retrieve foundational knowledge before specifics.
• Benefit: Highly contextual and can bridge complex vocabulary gaps.

Expands a query using the immediate conversational history or the agent's operational context. This is critical for stateful agents where a query is part of a longer dialogue or task sequence.

• Process: Maintain a context window of recent interactions. For a new query, append relevant entities, intents, or constraints from prior turns.
• Example:
  • User: "What's the revenue for Q1?"
  • Agent (later): "Compare it to last year."
  • Expanded Query: "Compare [Q1 revenue] to [previous year Q1 revenue]."
• Integration: Tied to agentic memory architectures and state management, ensuring coherence over extended interactions.

Used in multi-modal retrieval systems where a query in one modality (e.g., text) is expanded using concepts or features from another modality (e.g., image, audio).

• Process: A text query ("find sunny landscapes") might be used to generate a synthetic image embedding via a model like CLIP. This multi-modal embedding then queries a vector database containing both image and text embeddings.
• Application: Essential for agents with multi-modal memory encoding, allowing them to retrieve relevant memories regardless of the original data type.
• Foundation: Relies on aligned embedding spaces from models trained on paired multi-modal data.

Reranking is a two-stage retrieval process designed to improve precision. A broad set of candidate documents is first retrieved using a fast method (e.g., vector search on a bi-encoder), and then a more powerful, computationally expensive model re-scores them. A cross-encoder is the architecture typically used for this stage; it jointly processes the query and a single document through a transformer to produce a direct relevance score. Query expansion primarily aids the first-stage recall. The reranker then acts as a precision filter on these expanded results, using deep contextual understanding to identify the most relevant documents from the candidate pool.

Dense retrieval is a neural search paradigm where queries and documents are encoded into dense, low-dimensional vector embeddings (e.g., 768 dimensions). Relevance is determined by the similarity (e.g., cosine similarity) between these embeddings. Models like DPR (Dense Passage Retrieval) are trained specifically for this task. Query expansion can be integrated by generating an embedding for the expanded query text, or by using techniques like query encoding where the expansion terms influence the final query vector. This contrasts with sparse retrieval (e.g., BM25), where expansion explicitly adds term weights to a high-dimensional bag-of-words vector.

Knowledge graph retrieval involves querying a structured network of entities and their relationships to find factual information. Query expansion in this context often means traversing the graph to find related entities or concepts. For a query 'side effects of medication X', the system might expand by retrieving all entities connected to 'medication X' via a 'hasSideEffect' relationship. This provides a structured, deterministic form of expansion based on ontological relationships, complementing statistical or LLM-based term expansion. The retrieved sub-graph can then be converted to text and used as context in a RAG (Retrieval-Augmented Generation) pipeline.

Negative sampling is a critical training technique in contrastive learning for retrieval models (like those used in dense retrieval). During training, for a given query, non-relevant documents are selected as negative examples to help the model learn to distinguish relevant from irrelevant items. The quality of these negatives is crucial. Query expansion relates to the evaluation side of this problem: a well-expanded query should retrieve more of the true positive documents, thereby reducing false negatives in the retrieval results. Understanding the 'hard negatives' (irrelevant documents that are semantically close) is key to training models robust to the vocabulary gap.