Retrieval-as-memory hybrids

Modular graph-based RAG. Leiden community detection over knowledge graph; community summaries at multiple levels. DRIFT Search hybrid (Oct 2024).

Smaller classifier LM predicts query complexity, then routes to no-retrieval / single-step / iterative retrieval as appropriate. NAACL 2024.

Meta AI. Jointly fine-tunes retriever + LM for few-shot tasks.

"Don't Do RAG." Preloads all relevant documents into context, caches the resulting KV state; at inference the model answers from cache with no retrieval step — eliminates retrieval latency + retrieval-error failure modes.

Forward-Looking Active Retrieval. Uses model's low-confidence token predictions as signal to anticipate future information needs and retrieve proactively. Retrieves only when token probability falls below threshold; generator stays frozen.

Replaces chunk-retrieval with neuro-inspired generative memory. Operator maps observations to intermediate semantic structures; Reconciler integrates them into a persistent workspace with temporal/spatial/logical coherence. Targets 100k–…

Hippocampus-inspired RAG. LLMs + knowledge graphs + Personalized PageRank to mimic neocortex/hippocampus roles. NeurIPS 2024.

OSU follow-on to HippoRAG — adds online updates, recall-precision tradeoff calibration, and multi-hop reasoning improvements.

Multi-layer knowledge graph: higher layers hold coarse summary entities improving connectivity between similar lower-layer entities (HiIndex). HiRetrieval traverses across granularity levels.

LLM generates a hypothetical answer document; embeds that instead of the query. Closes the query-document semantic gap for retrieval. arxiv 2212.10496.

Uses model's generated output as a rich context signal for the next retrieval round rather than just the original query. Processes all retrieved knowledge as a whole, preserving generation flexibility.

Knowledge-Augmented Generation. LLM-friendly KG representation + mutual indexing between KGs and raw chunks + logical-form-guided hybrid reasoning. Handles numerical / temporal / multi-hop chains that vector similarity misses.

Performs minimal up-front indexing; defers LLM use until query time. Being integrated into the main GraphRAG library.

Hong Kong U graph-RAG variant — dual-level retrieval (local + global) over an entity-relation graph built incrementally from corpus updates.

Entity-relationship graph + broader-theme retrieval. Incremental update support. Integrates with RAG-Anything for multimodal docs.

UC Berkeley. Trains the model to ignore distractor documents and quote the relevant text in chain-of-thought.

Stanford recursive-summarisation tree over the corpus — clusters and summarises documents into a hierarchical index; retrieves at multiple levels. ICLR 2024.

Recursively embeds, clusters, and summarises chunks into a tree of abstractions. Retrieval traverses the tree at query time. ICLR 2024.

Foundational paper. Augments LM pre-training with a latent knowledge retriever — model retrieves and attends over Wikipedia-scale corpus during pre-training, fine-tuning, and inference. Backprop through retrieval over millions of documen…

DeepMind. Retrieval-Enhanced Transformer with chunked cross-attention over a 2T-token retrieval database. Frozen BERT retriever + differentiable encoder.

Frames RAG as Turing-complete with explicit monitored state variables, enabling adaptive retrieval halting and convergence control. Medical-LLM case study.

Applies AutoML to RAG: greedy algorithm evaluates and selects best-performing module at each pipeline stage (query expansion, retrieval, reranking, generation). Validated on ARAGOG (423 papers, 107 QAs).

Argues standard RAG breaks on bounded agent conversation streams (highly correlated, redundant spans). Decomposes the stream into semantic units, organises them into themes, then inverts that structure for retrieval.

Single embedding model supporting dense, sparse, and multi-vector (ColBERT-style) retrieval simultaneously. 100+ languages, inputs up to 8192 tokens. Self-knowledge distillation across the three modes.

Semantic chunking divides documents into coherent sections; LLM-based relevance scorer filters each chunk against the query before generation, pruning irrelevant chunks and reducing hallucination.

Lightweight retrieval evaluator scores docs and triggers different actions: use, supplement with web search, or discard. Decompose-then-recompose algorithm filters irrelevant content from retrieved passages.

Retrieval-augmented fine-tuning for knowledge graphs on graph databases.

Groups related Wikipedia documents into 4K-token units. Reduces NQ corpus 22M→600K units, HotpotQA 5M→500K. Hands longer passages to a long-context reader (GPT-4o).

RAG variant with a long-context global memory model that drafts answer clues used to guide passage retrieval — improves recall in noisy multi-hop QA.

Uses transformer multi-head attention activations (not decoder-layer embeddings) as retrieval keys, yielding multi-aspect embeddings that independently capture different semantic facets for heterogeneous queries.

Retrieves over an indexed graph by pruning candidate relational paths — reduces noise in graph-RAG vs flat path expansion.

Retrieves key relational paths from knowledge graph; flow-based pruning + reliability scoring; path-based prompting. Outperforms baselines on six datasets. arxiv Feb 2025.

End-to-end RL-trained RAG. Token-level policy chooses when to retrieve and which mode (text / KG / hybrid via RRF). Two-stage Group-Relative Policy Optimisation (GRPO). No public repo at time of writing.

Trains an LLM end-to-end to emit reflection tokens that decide whether to retrieve, assess passage relevance, and critique output quality — all within a single model.

Smaller distilled specialist generates multiple RAG drafts in parallel from distinct document subsets. Larger generalist verifies in a single pass. Subsets formed by content-similarity clustering to minimise redundancy.

Identifies optimal structure (table, catalog, algorithm, graph) per task; converts source documents into that structure at inference time and reasons over the structured form. Motivated by cognitive theories of human knowledge conversion.

Domain KGs + multi-stage web-page retrieval (sparse + dense) + LLM self-assessment of trustworthiness before finalising answer.