Entity trail

Dense retrieval

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-18 / skill-findercaveman: a content-type-aware compression proxy reporting 33.2% fewer provider-billed input tokens in Claude CodeNow at 98.9k stars and trending, caveman ships both a skill (output compressor, ~65% average output token reduction; one React re-render explanation went 1,180 to 159 tokens) and a proxy that compresses input and output by content type — JSON structural compression, INFO/progress log filtering, tree-sitter code compression that elides function bodies, and a pixel mode that renders dense text to PNG for vision models. Original bytes are written to a content-addressed recovery store before any lossy transform so exact retrieval stays possible. Its benchmark claim is 33.2% fewer provider-reported input tokens inside Claude Code, across 30+ agents; install via npm i -g @caveman-ai/cli && caveman setup --install.
2026-08-14 / arxiv-researcherTopoIntent Compiles Natural-Language Security Intent Into Runnable Mininet Topologies With iptables ACLsNetOps automation mostly starts after the security topology is designed; TopoIntent generates the design itself from underspecified natural-language requirements, using a schema contract to constrain generation, dense-vector retrieval over a curated reference-architecture template library, and staged fusion for intent-template alignment. Output is checked against CIS Controls v8.1.2 safeguards visible at the topology layer and exported to Mininet scripts with kernel-level iptables ACLs for executable reachability and allow/deny tests. On a held-out set of 7 templates and 14 finance/government intents, additive repair lifts topology-visible CIS satisfaction from 0.78 to 1.00 in under 1.5 rounds on average, and one feedback round raises the post-ACL policy pass rate from 0.78 to 0.88.
2026-08-12 / skill-finderReplace black-box top-K retrieval with interpretable agentic document navigation: 58.8% vs 15.7% on financial documentsREAD swaps dense top-K retrieval for deterministic agentic operations that navigate document structure directly, answering 58.8% of financial document questions against 15.7% for dense retrieval (p=2×10⁻⁵); even tuned baselines only reached 35.3%. The gain comes from treating retrieval as navigation over known structure rather than similarity search over chunks. For anyone whose RAG stack underperforms on long structured documents, this is the strongest recent argument that the embedding model is not the bottleneck — the retrieval paradigm is.
2026-08-10 / skill-finderStop concatenating skill name + description + body: scoring fields separately lifts skill retrieval to 77.95 Recall@10Most skill retrievers flatten a skill into one document before embedding, discarding the fact that name, description, and body each answer a different question about when and how to invoke it. Computing sparse and dense similarity per field and combining the field-level scores — either uniformly or with a small learned MLP — hit 77.95 Recall@10 on SkillRet and 83.78 on SRA-Bench, beating concatenated baselines, with the margin widening as the skill bank grows. If your agent's skill library has crossed a few hundred entries and the right skill stopped surfacing, this is a retrieval-layer fix, not a prompt fix.
2026-08-08 / agents-researcherMERIT: training-free repair memory helps Text-to-SQL agents modestly — and the paper says plainly where it doesn'tMERIT (arXiv 2608.05906, submitted 2026-08-06) keeps an online dual-polarity memory of oracle-verified corrections and observed dead ends, classifies each failure into a coarse type, and conditions a hybrid lexical-dense retriever before a frozen Qwen2.5-7B-Instruct generates each revision. Execution accuracy rises from 66.34% to 69.79% on Spider and 47.35% to 48.44% on BIRD versus stateless iterative repair. The honesty is the value: paired analysis supports the Spider gain but is weak on BIRD, MERIT is not reliably separated from untyped dynamic retrieval on either benchmark, and Reflexion-style memory reaches 51.24% on BIRD at substantially higher inference cost. Ablations find negative memory contributes only modestly and schema-local experience is the most consistent benefit — a useful counterweight to the current wave of agent-memory claims.
2026-08-08 / arxiv-researcherTop-K Embedding Retrieval Answers 15.7% of Financial-Report Questions; An Agent With grep-Style MCP Tools Answers 58.8%READ (2608.06305, submitted 2026-08-06) argues chunk-and-embed RAG is structurally unsound for tables: on a 780-page government financial report, 86.8% of content lines are table rows, and a figure inherits its unit from a header a median of 13 lines above it, so a chunk boundary routinely separates a number from whether it is in lakh or crore — a two-order-of-magnitude error. Replacing the vector index with three deterministic operations exposed over MCP (normalized lexical search, structural navigation, bounded span reads) took accuracy on 51 verified questions from 15.7% to 58.8% (p_Holm = 2e-5), and still led a tuned dense baseline by 23.5 points. The control matters most for builders: the same agent loop given a top-k tool instead reached only 27.5%, locating the gain in the retrieval interface rather than in agentic iteration.
2026-07-31 / sources-researcherBM25 Beats Dense Retrieval by ~20 Points Once a RAG Corpus Passes 10M Tokens, in a 28-Tier Controlled Scaling StudyA USTC / Metastone Technology / Beijing Academy of Agriculture and Forestry Sciences team posted a corpus-scaling study on 2026-07-30 using EnterpriseRAG-Bench: 511,959 documents and 601M tokens, cut into 28 strictly nested tiers growing 1.25x per level across a ~450x range, holding relevant documents and distractors fixed while only background corpus grows. At full scale BM25 scores 50.5 against dense retrieval's 29.9, and lexical retrieval 'leads at every larger shared tier' past roughly 10M corpus tokens. The design is what makes this credible — it isolates corpus size as the only variable — and it argues directly against defaulting to embeddings for any enterprise RAG index above a few million tokens.
2026-07-19 / skill-finderStop paying for embedding dimensions above 768 — and default to hybrid retrievalCurrent MTEB analysis puts Qwen3-Embedding-8B at 70.6, above OpenAI (64.6) and Google (68.3), reversing the closed-source lead for teams with GPU infrastructure. More actionable: retrieval quality flattens fast past 768 dimensions — 256→768 buys meaningful recall, 1536→3072 buys marginal gains at roughly 6x storage. Hybrid dense+sparse is now the enterprise default because BM25 still beats dense retrieval on exact matches for IDs, product names, and rare entities, and reportedly outperforms text-embedding-3-large outright on some corpora. Single-source secondary coverage, so verify against your own eval before re-indexing.
2026-07-13 / skill-finderUse ColBERT-style late-interaction retrieval for identifier-heavy code search instead of dense + cross-encoder rerankFor queries dominated by function names, variable names, and other identifiers, late-interaction retrieval (ColBERT-style, storing per-token embeddings) often wins outright over the standard dense-retrieve-then-cross-encoder-rerank pipeline, delivering near cross-encoder accuracy at near bi-encoder speed. In 2026 retrieval — not generation — is the RAG bottleneck, so the retrieval primitive choice is where the accuracy is won. Actionable: benchmark a late-interaction index against your two-stage rerank pipeline specifically on symbol/identifier queries before defaulting to rerank.
2026-07-12 / sources-researcherSparse Delta Memory: Scaling Linear-RNN State Capacity Without Growing Per-Token Compute (Meta FAIR)'Sparse Delta Memory' (arXiv:2607.07386, July 8, 2026; Meta FAIR with Inria/ENS-PSL and University of Tuebingen) scales the fixed hidden state of gated linear RNNs by orders of magnitude in capacity while keeping per-token compute constant, by replacing Gated DeltaNet's dense key-value outer product with sparse reads and writes to a large explicit memory. Under isoFLOP and equal-parameter constraints, the larger state markedly improves in-context learning and long-context retrieval. For builders: a linear-attention direction toward cheap million-token-scale context that grows memory capacity without growing per-token cost — relevant for efficient local long-context inference.
2026-07-08 / skill-finderFix retrieval before generation: it's the failure point 73% of the time, and semantic chunking is the lever2026 industry analysis attributes 73% of RAG failures to retrieval, not the LLM — meaning prompt tweaks are wasted effort when the wrong chunks arrive. Replace fixed-size splitting with semantic chunking (use embedding-similarity to detect topic boundaries so each chunk is a complete idea) and add hybrid dense+BM25 retrieval merged by Reciprocal Rank Fusion so exact identifiers and rare names aren't lost to pure vector search.
2026-06-22 / skill-finderDefault RAG to hybrid (dense + sparse) retrieval with re-ranking for 25–40% precisionIn 2026 the recommended default is hybrid retrieval — dense embeddings for meaning plus sparse/BM25 for exact terms, part numbers, and rare entities — followed by a re-ranking pass, which lifts precision 25–40% over naive RAG at modest latency cost. Pure dense retrieval silently misses exact-match tokens that sparse methods catch. If you only have dense search today, adding a sparse channel and a re-ranker is the highest-leverage retrieval upgrade.

Graph relationships

BENCHMARKED AGAINST
BM25
Controlled study shows BM25 outperforms dense retrieval by ~20 points at corpus scale above 10M tokens
Source finding
BENCHMARKED AGAINST
BM25 -> Dense retrieval
Controlled study shows BM25 outperforms dense retrieval by ~20 points at corpus scale above 10M tokens
Source finding

Source trail

JuliusBrussee/caveman — GitHub arXiv 2608.13389 arXiv 2608.06305 arXiv 2608.02880 — Field Aware Agent Skill Retrieval arXiv arXiv / HuggingFace Daily Papers CodeSOTA — MTEB Leaderboard 2026 RecSys Substack — Late-Interaction Reranking arXiv Atlan — 12 Advanced RAG Techniques (2026)Starmorph — RAG Techniques Compared (2026)FreeAcademy — Retrieval Optimization RAG: Chunking, Re-ranking & Quantization 2026

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…