Memory Hybrid Search

Sparse search, also known as lexical or keyword search, retrieves documents based on exact term matching. It uses algorithms like BM25 or TF-IDF to score documents by the frequency and rarity of query terms.

• BM25 is the modern standard, improving upon TF-IDF by factoring in document length normalization.
• Strengths: Excellent for precise keyword matching, proper nouns, and acronyms.
• Limitations: Fails with synonyms, paraphrasing, or semantic meaning (the 'vocabulary mismatch' problem).
• Example: A query for 'LLM fine-tuning' will miss documents mentioning 'large language model adaptation'.

Dense search uses neural embedding models to encode text into high-dimensional vectors. Retrieval is performed by finding stored vectors with the smallest cosine distance or Euclidean distance to the query vector.

• Embedding Models: Models like text-embedding-ada-002, BGE, or E5 convert text to vectors capturing semantic meaning.
• Approximate Nearest Neighbor (ANN): Indexes like HNSW, IVF, or PQ enable fast, approximate search over millions of vectors.
• Strengths: Excels at understanding intent, synonyms, and conceptual similarity.
• Limitations: Can perform poorly on exact keyword matches or rare technical terms not well-represented in the embedding model's training data.

This component merges the separate result lists from sparse and dense searches into a single, optimized ranking. Common fusion strategies include:

• Reciprocal Rank Fusion (RRF): A simple, effective method that combines rankings without needing relevance scores.
• Weighted Score Fusion: Assigns tunable weights (e.g., 0.4 sparse, 0.6 dense) to the normalized scores from each retriever.
• Cross-Encoder Rerankers: A final, computationally intensive step where a BERT-style model (e.g., Cohere Rerank, bge-reranker) scores the relevance of each candidate document to the query, providing the highest precision.

Acts as a pre- or post-retrieval filter to constrain results based on structured attributes, ensuring results are relevant not just in content but also in context.

• Common Filters: source, author, date_range, document_type, access_level.
• Implementation: Often applied using a vector database's native filter capabilities (e.g., filter=source == 'internal_wiki') during the ANN search.
• Role: Ensures operational compliance and precision by excluding irrelevant data scopes before semantic/keyword scoring occurs.

The subsystem that analyzes and potentially rewrites the user's raw query to improve retrieval performance for all downstream search components.

• Query Expansion: Adds synonyms or related terms (from a thesaurus or LLM) to improve sparse search recall.
• Hybrid Query Generation: Automatically decomposes a query into parts suitable for different retrievers (e.g., extracting keywords for BM25 and generating a semantic query for vector search).
• Spell Check & Correction: Corrects typos before they degrade retrieval accuracy.

The core semantic retrieval operation in a vector memory store. An agent finds the most similar stored embeddings to a query embedding using distance metrics like cosine similarity or Euclidean distance. This is accelerated by Approximate Nearest Neighbor (ANN) indexes such as HNSW or IVF, enabling fast similarity search across high-dimensional spaces. It forms the 'dense' component of hybrid search.

The preprocessing pipeline that prepares raw data for efficient hybrid search. Chunking algorithms segment documents into logical units (e.g., by paragraph, sentence, or token window). Semantic indexing then involves creating optimized metadata and embedding these chunks for retrieval. Techniques include:

• Recursive character splitting
• Semantic boundary detection using models
• Overlap strategies to preserve context
• Metadata enrichment for filtering

The end-to-end architecture where Memory Hybrid Search is typically deployed. A Retrieval-Augmented Generation (RAG) pipeline for agents includes:

1. Ingestion & Encoding: Documents are chunked and converted to vector embeddings.
2. Hybrid Retrieval: Combines vector and keyword search on the indexed data.
3. Ranking & Fusion: Results from different retrievers are scored and merged (e.g., using Reciprocal Rank Fusion).
4. Context Augmentation: Retrieved passages are formatted into the LLM's prompt.
5. Generation & Memory Update: The LLM generates a response, and the interaction may be stored.

A domain-specific language or API used to declaratively search an agent's memory. It unifies access across different search paradigms. Examples include:

• SQL for structured metadata filtering.
• Cypher for graph traversals in knowledge graphs.
• Vector search DSLs (e.g., vector_search('query', top_k=10)).
• Hybrid query syntax in databases like Weaviate or Pinecone, which allow combined statements like where { concepts: ["AI"] } AND nearVector({ vector: [0.1, 0.2] }).

A popular score fusion algorithm used to combine ranked results from different search techniques (e.g., vector and keyword). It computes a final score for each document based on its rank in each individual result set, without requiring the original relevance scores to be comparable. The formula is: score = sum(1 / (k + rank)) across all result lists. This makes it robust and effective for hybrid search, as it normalizes for the different score distributions of sparse and dense retrievers.

The underlying storage architecture principle for vector-based memory. Data is accessed by its content (or a content-derived key like a hash or embedding) rather than a fixed memory address. This enables associative recall. Implementations include:

• Vector databases: Query with an embedding to get similar embeddings.
• Hash tables: Key-value stores where the key is a hash of the content.
• Hopfield networks: A type of neural network that retrieves stored patterns from partial or noisy inputs.