Agent Know-How

Standard RAG treats historical support cases as static documents. RAFT abstracts each closed case into a directed chain of timeline entries and retrieves at the entry level, surfacing past cases whose intermediate states match the active one and returning the parent case's trajectory.

For troubleshooting agents, the retrieval unit should mirror how the work unfolds. Pick the index structure for the job, as with SQL over embeddings.

Cassie Shum's InfoQ talk lays out four patterns for building agents on a knowledge graph instead of basic RAG: context bundling, decision provenance, code as truth, and agent visibility. She demonstrates an engineering harness on a graph that shortens feedback loops and reduces token use.

Decision provenance puts attribution into the retrieval architecture itself rather than bolting it on as a separate scoring step.

KDDI, a major Japanese carrier, built Buffmee, a consumer RAG app that grounds answers in over 100 sources including books, magazines, and web media, on Google's Agent Development Kit. The design goal was high generation quality with fast responses across diverse media types.

At consumer scale the quality-versus-latency budget drives the retrieval architecture itself. See agent latency for the serving side.

A coding agent mid-task has already read much of what a retriever ranks highest, and a ranker can fill its budget with variants of one fact. MSS-Complement conditions on the agent's captured state and uses three semantic calls to propose a jointly sufficient set, find what is missing, and return 4-8 source units within budget.

On SERBench (500 held-out states from 45 repositories) it recovers a complete evidence set 73.0% of the time at five items, versus 61.4% for embedding-based reranking.

Databricks' retriever takes enterprise schemas and custom instructions as input and uses reinforcement learning (CISPO) to decide when one parallel search pass is enough and when a question needs sequential multi-hop search, with search cost in the reward.

It matches Claude Sonnet 5 and GPT-5.6 Luna answer quality at about 5.8s end-to-end, roughly half the latency, and the vendor reports it dominates their quality-vs-cost curves across the retrieval budget. Search effort becomes a learned per-query decision, not a fixed step count.

Tencent released WeMM-Embedding-9B on Hugging Face. Built on Qwen3.5, it embeds text, images, video, and visual documents into one 4,096-dimension space. It scores 80.6 average on MMEB-v2 (78 datasets), ahead of Qwen3-VL-Embedding's 77.8, and 59.5 on the harder MMEB-v3 (190 tasks).

A retriever that natively embeds slide decks and chart-heavy pages can reach multimodal evidence without a separate extraction pass.

REVA treats RAG compression as data mining: it aggregates a generator's historical query-document attention into a document-keyed, budget-agnostic score store, then renders plain-text views at any budget while keeping document order and the standard RAG interface.

The authors show existing compressors gain unstably over simple truncation and can add latency. REVA improves quality 1.0-5.8 points over them, cuts compression overhead 5.3x-15.6x, and adds under 40ms.

When retrieved documents mix useful and irrelevant context, models get distracted and hallucinate. SCoNE is a training-free edit: it finds context-aware FFN neurons with high attribution and high cross-input variability, then strengthens only those.

It needs a small mining sample, no fine-tuning, and no inference-time overhead, and reports consistent gains over baseline RAG methods on several knowledge-intensive QA benchmarks and two backbones. It applies to self-hosted open-weight models only.

AWS describes task-aware knowledge compression (TAKC) for analytical questions that span hundreds of documents, where plain RAG hits a ceiling. It pre-compresses the knowledge base into task-specific representations, caches them at multiple fidelity tiers, and routes each query to the right tier. An open-source implementation is included.

It trades an up-front compression pass for a smaller, denser context at answer time, instead of retrieving and re-reading raw pages per query.

Lazy Grounding shows a search agent can be misled without any false document. Using answer-changing rewrites of benchmark questions, each injected document truthfully supports a nearby question but surfaces for the original one, and agents adopt the nearby answer.

Accuracy drops 5.9 points on average and up to 17.3 points across 12 model-benchmark combinations. Fact-checking retrieved text will not catch it; the check has to be whether the evidence answers this question.

RAG systems trust whatever ranks as relevant, but relevance is not truth. This middleware combines NLI fact-checking, a five-signal poison detector, and a weighted Trust Index (0.4 factuality + 0.35 coherence + 0.25 (1 - poison)) with a dampener for heavily contaminated contexts.

• On TruthfulQA it reaches 91% accuracy and 100% recall on instruction-injection attempts.
• In-place edits such as entity swaps stay hard to catch.
• On FEVER, thresholds need per-model recalibration rather than transferring as-is.

See prompt injection for the attack side.

In multimodal RAG, attackers craft data whose embeddings align with benign entries in the vector space, so poisoned content is retrieved as if relevant. Existing query-time defenses (auxiliary detectors, similarity re-ranking, consistency checks) add inference overhead and generalize poorly. DSPrompt proposes a dynamic soft-prompt defense instead.

The attack targets the embedding space directly, so retriever ranking alone cannot be trusted as a relevance signal on untrusted corpora.

Multi-tool agents can state a true fact but attribute it to the wrong source, for example citing the account record for a fact that lives in the policy document. Pooled faithfulness scores miss this cross-source conflation.

ProvenanceGuard is a post-generation verifier that reads the captured tool trace with source IDs and never pools evidence. It splits the answer into claims, routes each to its likeliest source, checks support, compares that source with the one cited, and emits per-claim verdicts plus an allow/block decision, with no agent retraining.

For decisions that must be auditable (lending, fraud, clinical triage), ai-rete-rag runs two systems in series. A pure-Python Rete engine evaluates YAML rules: same facts, same verdict, every time. Then RAG retrieves passages from the organization's policy documents and an LLM writes a plain-English explanation citing them, with no ability to change the verdict.

The explanation cannot drift from what actually decided, the way a model's post-hoc rationale can. See agent reliability.

ActionRail, an open-source framework from ToolJet, adds runtime value and action grounding: before an agent's proposed action or value executes, it is checked against ground-truth business data.

Grounding moves from a retrieval-quality score after the fact to a deployable guard in front of side effects, aimed at the value-poisoning failure benchmarks measure (see agent benchmarks).