Shionogi's "Modular RAG" Boosts Search Accuracy from 50% to 90% — Whose Moat Is Agentic Search, Anyway?

---

TL;DR

1. Shionogi Pharmaceutical improved its internal search AI’s accuracy from 50% to 90% using “modular RAG.” However, the moat lies not in generic RAG technology but in the integration effort of structuring operational data, and the risk of platform absorption remains.
2. A paper on arXiv proposes injecting SSM hidden states to make RAG prefill O(1), speeding up prefill by up to 4,500x in edge environments.
3. OpenAI published message histories in a blog post to rebut Apple’s lawsuit. It reads as though what Apple is actually wary of is OpenAI’s hardware strategy itself.

Top Story: Shionogi Pharmaceutical Raises Search AI Accuracy from 50% to 90% with Modular RAG

Shionogi Pharmaceutical built an AI system with technology partner Classmethod to search pharmaceutical-development experiment plans and regulatory-compliant specifications, raising answer accuracy from 50% to 90%, according to their AWS Summit Japan 2026 presentation (Primary source, report). Because RAG using simple similarity search had capped accuracy at around 50%, Classmethod introduced a mechanism called “modular RAG.” The AI agent is given tools including an “OpenSearch tool” for cross-document search, a “file deep-read tool” for closely reading specific files, and an “Excel tool” for processing long-maintained operational Excel inventories, and it autonomously executes four steps: question analysis → search planning → search execution → answer generation. A loop is also built in that changes keywords or switches tools to search again if the agent judges the information to be insufficient.

The Technical Read

The main driver of the accuracy improvement is not the strength of the model itself, but the design that pre-classifies documents by type and use case, classifies question intent and retains context, uses tool selection and a re-search loop, and lets the agent read structured data — long-maintained operational Excel inventories — directly. What accounts for the jump from 50% to 90% is not a “smarter LLM” but a “searchable structure.” One caveat: this is a figure from a single implementation presented by the parties involved, and the public materials do not disclose the number of questions or evaluation criteria needed to judge reproducibility independently. It should therefore be interpreted cautiously. It also presupposes an environment where document classification and terminology dictionaries can be prepared in advance, and it remains uncertain whether this can be transplanted as-is to organizations lacking such preparation.

The Business Read

The design pattern itself — “give an agent multiple tools and have it autonomously build a search plan” — is technology heading toward being absorbed as a standard feature by tool-calling APIs from AWS Bedrock or OpenAI. Classmethod’s differentiator isn’t a generic algorithm; it’s the integration effort, including on-site interviews, of embedding itself in Shionogi’s operations to build out document classification, an internal terminology dictionary, and a knowledge base of business workflows. The moat sits upstream of the model — in Shionogi’s proprietary confidential data and business knowledge — while the modular RAG design pattern itself is reproducible by other systems integrators. Because terminology dictionaries and ledger organization need to be redone for each industry and company, packaging this into a resellable template is cost-heavy and hard to turn into a scalable “product.” On the other hand, highly regulated, highly confidential industries like pharmaceuticals are a gap that generic SaaS-style RAG struggles to penetrate, creating a market for specialized implementation work in regulated industries.

Implications and Positioning

If simple similarity search has plateaued, this case raises the priority of investing in tool decomposition, re-search, and terminology dictionaries rather than swapping models. Conversely, a strategy of selling the search algorithm itself as a differentiator carries a high risk of being swallowed by a platform’s standard features, and should be discarded as a hypothesis. For anyone considering a vertically specialized AI product, this case strengthens the hypothesis that what actually sells isn’t “smarter models or RAG,” but the integration effort of deciphering a customer’s confidential data structures and turning internal vocabulary and ledgers into a dictionary. Signal strength is medium, while confidence in the reproducibility of the 90% figure is medium-low: the design changes are concrete, but the evaluation conditions and number of questions are not public.

Other Notable Topics

PRECOG: SSM Hidden-State Injection Makes RAG Prefill O(1), Delivering a 4,500x Speedup at the Edge. A paper on arXiv proposes “PRECOG,” a method for State-Space Model (SSM)-based LLMs that, instead of stuffing retrieved documents directly into the context, pre-encodes the document corpus offline as SSM hidden states and injects the best-matching state directly at query time (Source). While a Transformer’s prefill cost scales with context length, SSM hidden states are fixed-size and position-independent, which the paper says allows prefill to be compressed to O(1). Experiments were run on a 1.2B-parameter gated-SSM model called “TENNs-LLM” (192KB hidden state), reducing prefill latency from about 27 seconds to under 6 milliseconds (roughly 4,500x) while maintaining answer quality comparable to standard in-context RAG. What matters technically is that this is something a Transformer’s KV cache cannot achieve in principle; on the business side, it means crossing the threshold from “long-document search on the edge is impractical” to “conversationally viable.” That said, this is a demonstration on a small, 1.2B-parameter model, and the paper’s abstract alone doesn’t reveal the cost of offline pre-encoding in real-world operation. So what: if you’re considering edge or offline use cases, it’s worth following SSM developments, but if you’re operating on a cloud-first basis where context length isn’t a bottleneck, this isn’t yet actionable.

Reddit Fights Back Against Suspected Brand-Linked “Stealth Posting” for AI Search Optimization. The Verge reported that companies are trying to place posts on Reddit that look like ordinary-user endorsements. In one skincare-focused subreddit with more than one million weekly visitors, a regular user flagged an account that repeatedly praised Honeydew Labs, and moderators removed it. No relationship between the account and the company has been established. Moderators then configured posts mentioning Honeydew Labs to be sent for review (Source). The article points out that Google has made Reddit pages more prominent in search results and that Reddit content is readily cited in AI search and chatbot answers, making this kind of posting economically worthwhile. Technically, this shows that the last line of defense is the human detection network of Reddit users and moderators. From a business standpoint, it’s evidence that a new market of “posting to get cited by AI search” is emerging, which for Reddit poses a risk of diluting the value of “authentic human voices.” Implication: growth tactics that rely on “astroturfed” posts on Reddit carry a high risk of being short-lived and backfiring once detection networks catch up, and this isn’t a tactic small-business founders should adopt. Owning primary source material — documentation, benchmarks, real data — that gets cited in AI search is a stronger medium- to long-term defense.

OpenAI Publishes Message Histories in a Blog Post to Rebut Apple’s Lawsuit. In response to a lawsuit Apple filed alleging trade secret theft by two former employees (Chang Liu and Tang Tan), OpenAI published iMessage and email histories in a blog post titled “Apple is getting this wrong” (Source). According to Reuters, Apple on Monday sought a preliminary injunction barring Liu, Tan, and OpenAI from using alleged confidential information. OpenAI said Apple’s preliminary-injunction request was “based on false information and completely unnecessary” because it neither possesses nor wants Apple’s trade secrets. Tan spent 25 years at Apple, where he led the design of the iPhone and Apple Watch, and is now OpenAI’s chief hardware officer. The public rebuttal via blog post carries more the flavor of managing perception in the talent market and public opinion than of settling the substantive merits of the lawsuit. Implication: more telling than the outcome of the lawsuit itself is the fact that Apple has begun treating OpenAI not as a “software AI company” but as a “hardware entrant.” If you’re planning AI-powered devices or wearables, you should build your positioning on the assumption that both companies are moving in earnest into adjacent territory.

IT Infrastructure Accounts for an Average of 67% of Healthcare AI Budgets (Wasabi Release). According to healthcare-sector results published by Wasabi Technologies from a survey conducted by the independent research firm Vanson Bourne, IT infrastructure accounted for an average of 67% of AI budgets among 171 healthcare respondents. They were part of a 1,700-person survey across 12 countries. The remaining 33% went to software including SaaS. The most commonly cited adoption challenge was storage cost and data access (48%), and 62% of organizations reported exceeding their cloud storage budget in 2025. Storage billing breaks down to an average of 50% for capacity-based charges and 49% for API operations and data transfer, among other items (Source). Because the storage vendor commissioned the survey, the conclusion that “storage cost is the biggest challenge” may reflect question design and sample selection favorable to its business and should be interpreted cautiously. Implication: estimating AI investment decisions based on token prices or model fees alone tends to underestimate the total cost, and it’s worth recalculating real costs including storage, data access, and transfer fees.

What to Try This Week / What Hype to Ignore

Worth trying: if your search or RAG accuracy has plateaued, it’s worth considering a design shift like Shionogi’s — tool decomposition, an autonomous re-search loop, and building out a terminology dictionary — rather than simply trying to make the model smarter. Worth ignoring: using “astroturfed” reviews on Reddit as an AI-search optimization tactic, since it’s short-lived against detection networks and cost, and doesn’t build a defensible asset. Worth holding off on: migrating cloud-based RAG to PRECOG’s SSM state injection right now, since it remains at the demonstration stage with a 1.2B-parameter model for edge use cases.