Retrieval Strategies & Workflows

Easy Knowledge Retriever provides multiple retrieval strategies to cater to different use cases, balancing between precision (specific details) and recall (broad themes), as well as between structured knowledge (Graph) and unstructured text (Vector).

Overview

Strategy	Mode Name	Description	Best For
Naive	naive	Classic Vector Search on text chunks.	Simple queries, exact matches in text.
Local	local	Entity-focused Graph Retrieval.	Specific questions about entities and their direct attributes.
Global	global	Relation-focused Graph Retrieval.	Thematic questions, understanding connections between concepts.
Hybrid	hybrid	Combines Local (Entity) and Global (Relation) Graph Retrieval.	Comprehensive Graph-based understanding.
Mix	mix	Combines Graph (Hybrid) and Vector (Naive) Retrieval.	The most robust strategy, leveraging both structured and unstructured data.
HybridMix	hybrid_mix	Combines Advanced Chunk Search (Vector + BM25 + RRF) and Knowledge Graph (Local + Global).	The most comprehensive strategy, utilizing all available retrieval methods.

Detailed Workflows

1. Naive Retrieval (naive)

Type: Vector Search (Unstructured)

This is the standard RAG approach. 1. Embedding: The user query is embedded using the EmbeddingService. 2. Vector Search: The system searches for the most similar text chunks in VectorStorage (Cosine Similarity). 3. Context Building: Retrieved chunks are concatenated to form the context. 4. Generation: LLM generates the answer based on chunks.

2. Local Retrieval (local)

Type: Graph Retrieval (Entity-Centric)

Focuses on specific entities mentioned in the query. 1. Keyword Extraction: LLM extracts "Low-Level" keywords (entities) from the query. 2. Entity Lookup: System looks up these entities in the Knowledge Graph. 3. Neighbor Retrieval: Retrieves direct neighbors (relations) of these entities. 4. Semantic Edge Filtering: Applies a Smart Re-ranking strategy to filter edges. * Constructs a local induced subgraph of retrieved entities. * Calculates local centrality (PageRank) to find structurally important nodes relative to the query anchors. * Prioritizes edges that connect important nodes or link two query entities directly. 5. Context Building: Constructs a context description from the matched entities and their filtered relations.

3. Global Retrieval (global)

Type: Graph Retrieval (Relation-Centric)

Focuses on broader relationships and themes. 1. Keyword Extraction: LLM extracts "High-Level" keywords from the query. 2. Relation Lookup: System searches for relationships in the Graph that match these keywords. 3. Context Building: Constructs a context description from the matched relationships and involved entities.

4. Hybrid Retrieval (hybrid)

Type: Graph Retrieval (Combined)

Combines Local and Global strategies. 1. Performs Local Retrieval to get entity details. 2. Performs Global Retrieval to get relational context. 3. Fusion: Merges both contexts to provide a comprehensive view of the structured knowledge.

5. Mix Retrieval (mix)

Type: Composite (Graph + Vector)

Combines the best of both worlds: Structured Graph knowledge and Unstructured Vector search. 1. Graph Retrieval: Performs Hybrid Retrieval (Local + Global). 2. Vector Retrieval: Performs Naive Retrieval (Vector Search on Chunks). 3. Fusion: Concatenates contexts from both sources. 4. Benefit: Handles queries where the answer might be in the graph structure OR in specific text details not captured by the graph.

6. HybridMix Retrieval (hybrid_mix)

Type: Ultimate Composite (Dense + Sparse + Graph)

Combines Advanced Chunk Search with Knowledge Graph retrieval. 1. Dense Search: Vector search on chunks (Semantic match). 2. Sparse Search: BM25 keyword search on chunks (Exact match). 3. RRF Fusion: Combines Dense and Sparse results using Reciprocal Rank Fusion. 4. Graph Retrieval: Performs Local (Entity) and Global (Relation) searches in the Knowledge Graph. 5. Benefit: Maximizes recall and precision by leveraging vector semantics, exact keyword matching, and structured graph knowledge simultaneously.

How to use

Pass the specific Retrieval class instance to the retrieval parameter when calling the query method.

from easy_knowledge_retriever.retrieval import MixRetrieval

# Example of querying with Mix strategy
response = await rag.query("What is the impact of X?", retrieval=MixRetrieval())

Or you can initialize the Retrieval classes directly if using them in a custom flow.