Ingest-Time Compilation Takes On Query-Time RAG, and Agentic Retrieval Meets Its Limits

A new arXiv paper argues for ingest-time semantic compilation over query-time interpretation in RAG, reporting 85.2% correctness on compiled claims versus 72.5% for chunk-based RAG. Another study introduces AgenticRAG-FP, showing that causal failure attribution in agentic RAG collapses to zero accuracy beyond the first hop. Additional research covers LineageRAG, which grounds GraphRAG evidence in verbatim source spans, and a benchmark revealing that agentic retrieval recovers performance on multi-hop queries over visually rich documents.

This digest covers RAG and GraphRAG developments from roughly August 17–24, 2026: new arXiv papers, Hugging Face Daily Papers picks, and posts from the blogs we track. 🔥 Highlights - RAG Deserves an Index: Why Ingest-Time Compilation Beats Query-Time Interpretation https://arxiv.org/abs/2608.20845v1 — pre-compile claims once, stop re-deriving meaning per query. - LineageRAG https://arxiv.org/abs/2608.16004v1 — grounds GraphRAG evidence in verbatim source spans, not vibes. - When Failures Propagate: Causal Failure Attribution in Agentic RAG https://arxiv.org/abs/2608.20627v1 — diagnosis collapses to zero past the first hop. - Temporal Validity on Real Software Histories https://arxiv.org/abs/2608.20685v1 — RAG serves stale facts over a third of the time. - Secure GraphRAG: Wiring Microsoft Copilot Studio to Neo4j with per-user Okta SSO https://neo4j.com/blog/graph-database/secure-graphrag-wiring-microsoft-copilot-studio-to-neo4j-with-per-user-okta-sso/ — per-user access control for enterprise GraphRAG. arXiv cs.CL / cs.AI - RAG Deserves an Index: Why Ingest-Time Compilation Beats Query-Time Interpretation https://arxiv.org/abs/2608.20845v1 2026-08-21 — Argues for "ingest-time semantic compilation": pre-computing atomic, provenance-validated claims at index time instead of re-deriving meaning from raw chunks on every query. Compiled claims hit 85.2% correctness on ~2.2k tokens vs. 72.5% on 16.3k tokens for chunk-based RAG. Practical value: a concrete architectural alternative if your pipeline is spending its context budget re-interpreting the same passages over and over. - Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate https://arxiv.org/abs/2608.16621v1 2026-08-17 — The engineering companion to the paper above: shows incremental low-rank SVD updates to a semantic index are 33.7x cheaper per update than a full rebuild, while tracking the rebuilt index to floating-point precision. Answers the obvious objection to ingest-time compilation "won't re-indexing kill me on updates?" with a working incremental scheme. - LineageRAG: Harnessing GraphRAG by Constructing Evidence Lineages with Source Grounding https://arxiv.org/abs/2608.16004v1 2026-08-17 — Builds explicit "evidence lineages" per query, grounding each hop of graph traversal in a verbatim source span instead of leaving the evidence trail implicit. Beats leading GraphRAG baselines by 3.51 R@5 / 5.96 EM / 5.22 F1 on HotpotQA, 2WikiMultiHopQA, and MuSiQue — useful if your GraphRAG answers are accurate but impossible to audit. - When Is Complex Chunking Worth It? A Multi-Objective Evaluation of Chunking Methods at Scale https://arxiv.org/abs/2608.16586v1 2026-08-17 — Evaluates 8 chunking strategies across 2 corpora, 3 embedding models, and multiple corpus sizes on both retrieval quality and operational cost throughput/latency/memory . Bottom line: complex chunking rarely pays off consistently once you account for cost — good ammunition against over-engineering your chunker. - When Failures Propagate: Causal Failure Attribution in Agentic RAG https://arxiv.org/abs/2608.20627v1 2026-08-20 — Introduces AgenticRAG-FP, which injects a certified fault at a specific hop and checks whether post-hoc diagnosis can still find it once the agent's trajectory has changed. Coverage-based diagnosis collapses to 0.00 accuracy beyond hop 1 — a hard warning for anyone shipping "explain why the agent failed" tooling on top of agentic RAG. - VisDocAgentBench: Benchmarking Agents for Visually Rich Document Retrieval https://arxiv.org/abs/2608.17889v1 2026-08-18 — A strong static retriever hits 97.5% R@1 on direct queries over visually rich documents but only 2.5% on two-bridge multi-hop queries; agentic search recovers most of that loss. Quantifies exactly where the extra cost of agentic retrieval is earned, and where it's wasted. - Temporal Validity on Real Software Histories: Eliminating Stale-Fact Errors in Code-Assistant Memory https://arxiv.org/abs/2608.20685v1 2026-08-21 — Validated on 130 real GitHub-issue state transitions: standard RAG has no time model and serves superseded facts 36–38% of the time, even with an LLM reranker in the loop. A deterministic supersession-memory system MemStrata drives that down to ~0 at comparable latency — sharp, data-backed illustration of a known RAG failure mode. - Weighted Memory Tree: Remembering What Matters for Long-Horizon LLM Agents https://arxiv.org/abs/2608.20631v1 2026-08-21 — A hierarchical memory tasks/subtasks/actions with per-memory retention scores and decay, instead of a flat growing transcript. Gains +9.97 points accuracy on GAIA-Text over linear memory while cutting prompt tokens by 32.8% — a practical pruning technique for agents that run long. Hugging Face Daily Papers - Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence https://huggingface.co/papers/2608.21156 2026-08-21 — A position paper proposing "Graph Engineering" as the successor to prompt/context/harness engineering: explicit, evolving graph structures coordinating tasks, agents, and system state across multi-agent systems. Worth reading if you think of GraphRAG purely as a retrieval technique — this argues the graph itself becomes the coordination layer for the whole agent system. - EviRank: Structured Relevance Evidence for Multimodal Image Re-ranking https://huggingface.co/papers/2608.20886 2026-08-21 — Recasts reranking as constraint satisfaction: parses queries into typed criteria required/forbidden/ignorable across six semantic slots, then reranks via deterministic rubric scoring plus evidence-grounded listwise comparison. State of the art across five multimodal retrieval benchmarks, and a distilled student model retains 90%+ of teacher quality at much lower cost — relevant if your reranker's failures are hard to explain to stakeholders. - MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use https://huggingface.co/papers/2608.20202 2026-08-20 — Shows that even faithfully retrieved, relevant memories can distort reasoning reasoning fixation, belief distortion : every tested memory strategy in the benchmark underperforms a no-memory baseline, with the best methods still losing more than 10%. Proposes a simple inference-time instruction fix AdaptiveMem . A useful reality check before you ship an agent-memory feature on the assumption that "more retrieved context is strictly better." Anthropic Engineering Blog Nothing relevant this week — no new posts on retrieval, agents, or memory engineering in the window. LangChain blog Nothing squarely on-topic this week. LangSmith shipped "Tuned Evaluators" 2026-08-18 for flagging perceived error in agent traces, but it's general agent-observability tooling rather than a retrieval-quality or RAG-specific development. LlamaIndex blog - How LlamaIndex Uses Temporal to Scale Reliable Document Orchestration https://www.llamaindex.ai/blog/temporal-scale-document-orchestration 2026-08-17 — Describes migrating their document-processing pipeline from RabbitMQ to Temporal for durable workflow execution at millions-of-pages/day scale, addressing fairness and concurrency in ingestion. It's about ingestion infrastructure rather than retrieval quality, but directly relevant if your RAG pipeline's bottleneck is reliably getting documents indexed in the first place. Neo4j blog - Secure GraphRAG: Wiring Microsoft Copilot Studio to Neo4j with per-user Okta SSO https://neo4j.com/blog/graph-database/secure-graphrag-wiring-microsoft-copilot-studio-to-neo4j-with-per-user-okta-sso/ 2026-08-20 — Walks through wiring Microsoft Copilot Studio to a Neo4j-backed GraphRAG system via a Neo4j MCP server on Cloud Run, forwarding each user's Okta access token so graph queries respect per-user permissions instead of a shared service account. Solves the "confused deputy" problem for agentic GraphRAG in enterprise settings that need row/entity-level access control. - This Week in Neo4j: Agent Memory, MCP, Skills, Cypher and more https://neo4j.com/blog/twin4j/this-week-in-neo4j-agent-memory-mcp-skills-cypher-and-more/ 2026-08-21 — Roundup pointing to real developer work in-window: Aura instances now ship a built-in MCP server for zero-setup OAuth graph access from clients like Claude Desktop, and skill distillation in Neo4j's Agent Memory Service converts agent memory into portable SKILL packages. Useful as a pointer list if you're evaluating Neo4j's agent-memory tooling specifically. Simon Willison - ChatGPT search now uses the site: operator at scale https://simonwillison.net/2026/Aug/20/chatgpt-search-now-uses-the-siteoperator-at-scale/ 2026-08-20 — Citing GEO-monitoring firm Promptwatch, ChatGPT Search's use of the site: operator jumped from 0.3–0.5% to 16–17% of queries after the GPT-5.6 rollout, alongside a drop in Reddit citations. Simon notes this is likely exposed via a structured search query, recency, domains function rather than a raw operator string — a concrete data point on a production LLM system shifting toward explicit domain-restricted retrieval to improve source reliability. Latent Space Nothing relevant this week — posts in the window covered agent harness design, behavioral simulation, and model routing economics, none with substantive retrieval or memory-architecture content. Interconnects Nothing relevant this week — the one post in the window covered open-model business strategy, not retrieval or evaluation. Through-line The strongest thread this week is a shift from treating retrieval as a query-time problem to treating it as an ingest-time and provenance problem: pre-compiled semantic indexes, evidence lineages grounded in verbatim spans, and Neo4j's per-user access wiring all move work and accountability earlier in the pipeline instead of hoping the query-time retriever gets it right. The second thread is a set of sober benchmarks pushing back on assumptions: agentic retrieval only earns its cost on genuinely multi-hop queries, more retrieved memory can actively hurt reasoning, and plain RAG has no notion of time and will confidently hand back superseded facts more than a third of the time. Taken together, it's a good week for treating "retrieval works" as a claim to verify per failure mode, not a property you get for free. What's your pipeline doing about stale facts or evidence provenance? Drop a comment below.