Context retrieval became a product category on November 5. At BoxWorks 2026 in San Francisco, Box announced Box Context Retrieval, an engine built on NVIDIA Nemotron Retriever models that turns enterprise files into a structured knowledge layer for agents. Box says customers will have the potential to improve search accuracy by 20 percent on Recall@5, reduce token usage by 90 percent, and accelerate question answering fourfold. The engine will be generally available to Enterprise Advanced customers early next year. NVIDIA Nemotron models, beginning with Nemotron 3 Ultra, are available now in Box AI to the same tier. Source: Box via Business Wire, syndicated at https://lifestyle.folsomlocalnews.com/story/987519/box-partners-with-nvidia-to-deliver-the-trusted-content-layer-for-enterprise-ai/ and reported by SiliconANGLE at https://siliconangle.com/2026/11/05/box-taps-nvidias-ai-models-to-build-a-rich-context-layer-for-autonomous-enterprise-agents/ and by Reuters via SRN News at https://srnnews.com/box-expands-nvidia-partnership-unveils-ai-workflow-builder/

The numbers are the lede, but they are not the story. The story is that retrieval, the step that decides which passages an agent sees before it answers, is now being sold as a separate layer with its own models, its own benchmark, and its own price tier. A passage is cheaper than a file. It is also weaker than a record. This article sets out what Box announced, why three other vendors made the same move in the same week, what the independent evidence says about token cost and accuracy, where the counter argument is strongest, and the test a retrieved passage has to pass before it should govern work.

What Box actually announced

Box describes the engine in its release as one that transforms enterprise data into a structured knowledge layer that agents can reason over. With Nemotron Retriever embedding models, it will retrieve permission aware information across millions of files, including scanned documents, images, audio, and video. Permission aware is defined in the release as providing agents with information from only the content users are permitted to access. The release adds faster indexing, so new and updated content becomes available to agents in real time, and hybrid search, which combines keyword matching with an understanding of meaning.

Box chief executive Aaron Levie said: "Enterprise AI reaches its full potential when models and agents can securely access the unique context of a business." He added that the partnership creates "a new intelligence layer for enterprise content." NVIDIA founder and chief executive Jensen Huang said: "AI agents need more than intelligence. They need the knowledge and context of the business to do useful work." SiliconANGLE reported the same launch under the name Box Content Retrieval, noted that agents adhere to native file permissions for each piece of content, and placed the engine alongside Box Architect, a workflow builder entering beta later this year, and new threat detection for rogue agents. Availability matters for how to read the figures. The retrieval engine is not generally available until early next year, so the 20 percent, 90 percent, and fourfold figures are vendor projections for a product customers cannot yet run at scale, not measured results from production use. Box gives no baseline dataset in the release, and it does not state whether the 20 percent is relative or in percentage points beyond naming Recall@5.

Why context retrieval split off this week

Box was not alone. On November 4, PhoenixAI launched Contextbase at its NEXT QUERY 2026 summit. Contextbase stores semantics, knowledge, memory, traces, and evaluations as context in one system. Its announcement states the problem plainly: definitions live in a semantic layer, documents in a vector store, memory in a memory service, and traces and evaluations in observability tools, and each keeps its own copy and its own permissions. Contextbase holds typed, versioned knowledge records, such as a spec, an incident review, or a playbook, linked to each other, alongside semantic views defined in Semantic SQL, memory that agents write from earlier work, traces of every model and tool call with prompts, outputs, tokens and cost, and evaluations from test datasets and live traffic. It runs in private preview from November 4. PhoenixAI co founder and chief executive James Li said: "An agent is only as good as its context, and today it has to glue that context together from separate systems on every turn. What it gets is partial, inconsistent and often stale." Source: https://www.phoenixdata.ai/newsroom/phoenixai-launches-contextbase-the-context-infrastructure-for-ai-agents

On November 5, Eon launched Eon Yukon, a data layer for agents that builds context during ingestion, including statistics, changes, and descriptions informed by business definitions, and makes that context available through the Model Context Protocol. Eon states that fresh data alone does not tell an agent what a column means or what has changed. In its own ingestion benchmarks, Yukon used between 4 and 7 times less compute than Apache Spark and Flink while maintaining near real time freshness, and in a replay of one hour of production traffic from Rise, an advertising platform, it held a 60 second freshness target using 85 percent fewer machines than Flink and more than 88 percent fewer than Spark. Those are vendor benchmarks, stated as such here. Source: https://www.globenewswire.com/news-release/2026/11/05/3396659/0/en/eon-launches-high-performance-ai-data-layer-with-less-infrastructure.html

Three launches in two days share one structure. Each vendor separates the context layer from the model. Each treats meaning, freshness, and permission as properties to build at ingestion or retrieval time, not as prompts to write at query time. Each sells the saving in tokens, compute, or machines. The convergence is the evidence that the bottleneck has moved. The model is no longer the scarce part. The passage the model receives is.

The evidence on tokens and accuracy

The strongest independent comparison in this space does not come from a vendor. Hamilton and colleagues, in a June 2026 preprint discussed by HackerNoon, ran long context prompting against semantic retrieval on the same document grounded task. Long context achieved the highest correctness, 73.1 percent against 65.4 percent for semantic RAG, but at 26 times the per query token cost. The arithmetic in that discussion is illustrative and worth repeating with its assumptions visible. A 200,000 token prompt at an illustrative 3 dollars per million input tokens costs 0.60 dollars per query before any output. At 50,000 queries a month, that is 30,000 dollars. The same question answered from six retrieved chunks, about 4,000 tokens, costs about 0.012 dollars per query, or about 600 dollars a month. Prompt caching narrows the gap when the prefix is identical across queries. It does not narrow it when the prefix differs by user, by tenant, or by hour, which is the normal enterprise case, because permissions and freshness differ by reader. Source: https://hackernoon.com/long-context-beats-rag-on-accuracyso-why-keep-retrieval

Scale changes the ranking again. Wang and colleagues, in a scaling study posted on July 30, 2026, varied corpus size across 28 nested tiers from 1,144 to 511,959 documents, 1.7 million to 601 million tokens, on EnterpriseRAG-Bench, with 500 questions and a fixed set of relevant and adversarial documents. A file system agent that explores the corpus with tool calls led at the smallest tiers, but its sequential exploration cost 39 times more query tokens at the bedrock tier. Around 10 million corpus tokens, lexical BM25 retrieval overtook it and led at every larger shared tier, with a margin approaching 20 points at full scale. Dense retrieval remained efficient but less accurate, and graph based methods reached construction limits before deployment scale in that study. Source: https://arxiv.org/html/2607.26497

Read together, the two studies explain Box's commercial bet. At small, stable, single audience scale, putting material in the window can win on correctness. At enterprise scale, with millions of files, per user permissions, and content that changes during the day, retrieval is not optional, and the quality of the retriever sets the ceiling for every answer that follows. A retriever that never surfaces the right passage caps accuracy at zero for that question, no matter how strong the model is. That is why a 20 percent move in Recall@5 matters more than a model release for a buyer whose corpus lives in files. It is also why hybrid search, which Box names in its release, has become the default design in the research: keyword matching carries exact identifiers, product codes, and figures, where embeddings are weakest, while dense scoring carries paraphrase and concept matches.

The counter argument, stated fairly

The counter argument has three parts, and each part is correct on its own terms.

First, long context beat retrieval on correctness in the Hamilton comparison, by nearly eight points. Any claim that retrieval is more accurate in general argues against a measurement. Retrieval buys cost control, freshness, and permission enforcement. On a corpus that fits in the window, it does not buy peak accuracy.

Second, a long window is not a uniform window. The same HackerNoon discussion cites Chroma's Context Rot study from July 2025, which evaluated 18 models and found performance varies significantly as input length changes, with degradation worsening when distractors are present, and NoLiMa, presented at ICML 2025, which tested 13 models claiming 128,000 tokens or more and reported that at 32,000 tokens, 11 models dropped below 50 percent of their short length baselines. Those results cut against both camps. They weaken the case for stuffing a corpus into a prompt, and they also warn that a retriever returning more passages is not returning better context.

Third, Box's figures are projections for a product that reaches general availability early next year. No independent replication, no published test corpus, and no customer production result accompanied the announcement. The same caution applied in this blog to Perplexity Brain, to Salesforce Agentforce Memory, and to every vendor evaluation cited here. A projection is a hypothesis with a date on it. The date to test it is general availability, on the buyer's own labelled questions, not the launch keynote.

None of those points restores the old default. The old default was a fixed chunk pipeline, one embedding model, top five results by cosine similarity, and no re ranking, no permission check at query time, and no record of which version of a passage the agent saw. Long context made that pipeline obsolete. The launches this week are attempts to replace it with something measurable.

The Retrieved Passage Test

A retrieved passage solves the token problem. It does not solve the record problem. A passage can be relevant, permission aware, and cheap, and still be the wrong version, superseded last week, drawn from a draft, or contradicted by a later decision in another system. When an agent acts on a passage, the company needs to answer five questions about that passage later, in an audit, in a dispute, or in a correction. Jilu calls this the Retrieved Passage Test.

One, source. Name the file, system, and owner the passage came from, not only the index that returned it. Two, date and superseded version. State when the source was written, when it entered the index, and which earlier version it replaced, so two passages that disagree can be ordered instead of averaged. Three, permission at query time. Record the permissions of the person who asked, checked when the question was asked, not when the file was indexed, because access changes between those two moments. Four, scope of correction. When the source is corrected, state which summaries, skills, and queued actions built on the old passage are withdrawn or rebuilt. A correction that updates the file but leaves the derived answer in place has not propagated. Five, receipt. Save the exact passages the agent saw for that answer, with their source and date, as a receipt that can be reopened. Without a receipt, a team cannot tell whether the agent failed to retrieve, retrieved the wrong version, or reasoned badly from the right passage, which is the same distinction Salesforce AI Research raised this week for agent memory.

Box Context Retrieval, on its own release, addresses permission aware retrieval, real time indexing, and multimodal coverage. PhoenixAI Contextbase, on its own announcement, addresses versioned records, traces, and evaluations in one store. Neither announcement, on its own text, claims the full test: correction propagation to derived work, and a saved receipt per answer that a person can inspect and dispute. That gap is not a criticism of either launch. It is the boundary between a retrieval layer and a company record. Retrieval decides what the agent sees. The record decides what the company stands behind, who approved it, and how a correction reaches every place the old fact went.

What to ask before buying context retrieval

Buyers evaluating any context retrieval engine, including Box's, can settle the purchase with their own numbers in an afternoon. Run Recall@5 and Recall@10 on at least 200 labelled questions drawn from real work, including questions whose answers changed in the last 90 days and questions the asker is not permitted to answer in full. Measure tokens per answered question and cost per correct answer, not cost per query, because a cheap wrong answer is not a saving. Test a correction end to end: change a policy in the source, then ask a dependent question in a fresh session and check whether the new version governs and the old passage is marked superseded. Test a refusal: ask for a composite answer built from several permitted files that together disclose a restricted conclusion, and check whether the system scores the answer, not only the retrieval. Save every failed case. Those failed cases are the specification for the record layer that retrieval alone does not provide.

The launches this week will be judged, correctly, on recall, token cost, and speed when they reach customers. Those are retrieval numbers. The work that decides whether an enterprise trusts the answer sits one layer behind them, in the record: source, date, superseded version, approver, and a correction path that reaches the answer after the file changes. Context retrieval makes the passage cheap enough to use at scale. The record makes it safe enough to act on. Companies need both, and they should not mistake the first purchase for the second.