CodaRAG: Connecting the Dots with Associativity Inspired by Complementary Learning

1. Problem Statement

Large Language Models (LLMs) struggle with knowledge-intensive tasks due to hallucinations and fragmented reasoning, particularly when information is dispersed across multiple sources. While Retrieval-Augmented Generation (RAG) mitigates this by grounding generation in external evidence, existing methods suffer from two primary limitations:

• Flat Context Construction: Naive RAG treats retrieved documents as isolated chunks, failing to reconstruct the logical chains connecting them.
• Ineffective Graph Utilization: Existing GraphRAG methods (e.g., LightRAG, HippoRAG) often rely on local one-hop expansions or simple relevance scoring. They retrieve "dots" (entities) but fail to "connect the dots" into a coherent evidence chain, leading to fragmented contexts that hinder complex reasoning.

The core challenge is the absence of a controlled associative retrieval mechanism that can simultaneously expand to uncover distant associations, organize evidence into a coherent structure, and suppress interfering noise.

2. Methodology: The CodaRAG Framework

Inspired by Complementary Learning Systems (CLS) theory—which posits a synergy between a fast, associative hippocampus and a slow, integrative neocortex—CodaRAG transforms retrieval from a passive lookup into an active, multi-stage associative discovery process. The framework operates via a three-stage pipeline:

Stage I: Knowledge Consolidation (Identifying the Dots)

This stage unifies fragmented extractions into a stable memory substrate (a consolidated Knowledge Graph, KG).

• Entity Type Discovery: Unlike rigid schemas (ClosedIE) or untyped extraction (OpenIE), CodaRAG induces semantic categories directly from corpus-level patterns. This ensures consistent identification of semantically similar entities (e.g., aligning "Google," "Google Inc.," and "Alphabet" under a unified type).
• Fragmented Entity Merging: To address entity fragmentation across different documents, the system uses embedding similarity to identify candidate pairs, followed by an LLM-based judge to merge them. This reduces redundancy and repairs structural gaps, creating a canonical memory foundation.

Stage II: Associative Navigation (Connecting the Dots)

Instead of relying on a single retrieval metric, CodaRAG employs three complementary pathways to traverse the graph, mimicking human spreading activation:

1. Semantic Association (Fast/Local): Expands locally from entry entities based on semantic similarity between query cues and relations/entities. It retrieves explicitly related, topologically close evidence.
2. Contextualized Association (Slow/Global): Uses Personalized PageRank (PPR) to propagate query relevance globally across the graph. This identifies salient concepts and organizes dispersed evidence into a coherent subgraph, even if they are not directly connected to the query.
3. Functional Association (Structural): Identifies structurally analogous entities using Fast Random Projection (FastRP) embeddings. This captures higher-level relational patterns (e.g., finding a "Samsung" when the query is about "Apple" based on functional market roles), bridging gaps that semantic similarity misses.

Stage III: Interference Elimination (Refining the Context)

To prevent "hyper-associativity" (where loose connections introduce noise), an executive control mechanism filters the retrieved subgraph.

• Noise Pruning: An LLM evaluates the retrieved entities and relations against the query, removing associations that do not meaningfully contribute to the answer.
• Context Construction: The remaining high-precision evidence is aligned with source text chunks to construct the final context for generation.

3. Key Contributions

• CLS-Inspired Architecture: The first RAG framework to explicitly model the dual-process nature of human memory (fast association vs. slow integration) to solve evidence fragmentation.
• Multi-Dimensional Navigation: Introduces a novel tri-pathway retrieval strategy (Semantic, Contextualized, Functional) that captures diverse relational patterns beyond simple vector similarity.
• Controlled Associativity: Proposes an "Interference Elimination" stage that acts as a cognitive filter, balancing the trade-off between broad evidence coverage and reasoning precision.
• Unified Memory Substrate: Develops a robust indexing pipeline that consolidates fragmented entity mentions into a canonical KG, solving the "entity linking" problem inherent in dynamic RAG systems.

4. Experimental Results

The framework was evaluated on GraphRAG-Bench across two datasets (Medical and Novel) and four task types (Fact, Reasoning, Summarization, Creative Generation).

• Performance Gains: CodaRAG achieved absolute gains of 7–10% in retrieval recall and 3–11% in generation accuracy compared to strong baselines like HippoRAG 2 and LightRAG.
• Task-Specific Improvements:
  • Reasoning & Facts: Significant improvements in accuracy due to better logical chain reconstruction.
  • Summarization: Demonstrated a superior coverage-accuracy trade-off, maintaining high accuracy while covering broader evidence than baselines.
  • Creative Generation: Produced more grounded and faithful responses by synthesizing dispersed evidence effectively.
• Ablation Studies: Confirmed that all three stages are critical. Removing any association pathway degraded performance, and removing Interference Elimination led to a drop in faithfulness due to noise.
• Robustness: The method remained effective when the backbone LLM was switched from GPT-5-mini to Mistral-medium-3.1, proving the approach is model-agnostic.

5. Significance

CodaRAG represents a paradigm shift in RAG systems, moving from retrieval-as-lookup to retrieval-as-reasoning. By treating retrieval as an active process of connecting disparate pieces of information through associative pathways, it addresses the fundamental limitation of current LLMs in handling complex, multi-hop queries. The framework demonstrates that structured, interference-aware retrieval is essential for overcoming evidence fragmentation, offering a scalable solution for high-stakes domains like healthcare and historical analysis where precision and logical coherence are paramount.