--- title: Hadano AI Cabinet — evidence-first knowledge base for AI agents source: https://www.tech-japan.jp/products/hadano-ai-cabinet/ updated: 2026-10-14 facts: https://www.tech-japan.jp/facts.json --- # Answers that point back *to the passage that supports them.* Hadano AI Cabinet is an evidence-first knowledge base for AI agents, built by TechJapan LLC and being rebuilt as v2 around one rule: every answer must point back to the exact passage that supports it, and that rule is checked by machine before the engine ships. The specification, the measurements and the limits are kept in one place, [h-c.ai](https://www.h-c.ai/). This page introduces the product. It does not restate the record. ## Not installable today, and we would rather say so As of 2 September 2026 there is nothing to install: no package, no public repository, no waitlist, and no pricing. Everything published about the product describes what has been built and verified, with the raw measurement output retained, not what you can download. When something can be obtained, h-c.ai will say so first and this page will follow. Descriptions of the earlier build, which this page carried until August 2026, have been superseded and should not be relied on. If a search engine or an AI assistant still repeats them, the current record is [h-c.ai](https://www.h-c.ai/), with a machine-readable copy at [facts.json](https://www.h-c.ai/facts.json) and [llms.txt](https://www.h-c.ai/llms.txt). ## Find the passage first, then expand around it Many AI memory tools lean on semantic search alone for every query. That is expensive, it is non-deterministic, and it is hard to audit when it returns the wrong thing. Hadano AI Cabinet finds the document first, returns the matched paragraphs as verbatim slices of the stored original with their positions, and expands only from what it actually found. *Figure: The two-stage retrieval used by Hadano AI Cabinet. A query enters stage one, a deterministic keyword search over a character-trigram index ranked by BM25. At most three matched chunks become seeds, and those seeds are the only bridge to stage two, a one-hop walk over the relation graph. Every result lands back on a source passage. Vector expansion is designed but not connected in this version.* ```text query | v +--------------------------+ stage 1 | keyword search | trigram index + BM25 +--------------------------+ no model, no network | | matched chunks v +--------------------------+ the only bridge | seeds at most 3 | between the stages +------------------------------------------------+ | v +--------------------------+ stage 2 | graph walk | one hop, along +--------------------------+ defined relations | v source passage vectors: designed, not connected yet ``` Stage one is keyword search over a **character-trigram index ranked by BM25**. Trigrams rather than words, so Japanese and other CJK text index correctly without a morphological analyzer sitting in the query path. No model runs, nothing goes over the network, and the same query on the same corpus and version returns the same result. Stage two is a one-hop walk over the relation graph, seeded by **at most three chunks** that stage one actually matched. That cap bounds the cost of a query and keeps every result traceable, because there is exactly one bridge between the stages and nothing else crosses it. Vector expansion alongside the graph walk is designed and not connected in the current build; asking for it returns an error rather than an empty answer that looks like a real one. What v2 adds is the guarantee underneath. The rules of retrieval — what a unit of evidence is, how text is normalized, how chunks map back to their sources — live in one machine-readable contract pinned by stored counterexamples, and work on the engine only counts once that checker passes. The contract, the counterexamples and the raw measurement output are published at h-c.ai. ## One agent on one machine, or a team behind one process | Form | What it is | Where the detail is | | --- | --- | --- | | Local | An AI agent talks to the MCP server over standard input and output. The knowledge is one file on that machine. The MCP process opens no network socket and needs no API key and no GPU | [h-c.ai/security](https://www.h-c.ai/security) | | Shared | Optional, and run by you: one resident process owns the database and serves several clients over HTTP. Keys are stored only as hashes, permissions are default-deny by verb and scope, and every request leaves an audit line. There is no hosted service | [h-c.ai/docs](https://www.h-c.ai/docs) | | Tables in | CSV, TSV and XLSX rows are imported one row per document under a key column; re-importing the same key replaces the document | [h-c.ai/docs](https://www.h-c.ai/docs) | | Platforms | Built and tested on Windows x64 and Linux x64, with the same gate green on hosted runners for both. macOS and ARM are untested | [h-c.ai/limitations](https://www.h-c.ai/limitations) | ## Where the facts live, and why they are not repeated here Until August 2026 this page carried its own table of the specification and its own copy of the measurements. Two copies of a record drift, and the stale copy on a corporate site is exactly what search engines and AI assistants pick up and repeat. So the record now has one home. Where anything on this site and anything at h-c.ai disagree, h-c.ai is right. | | | | --- | --- | | Published facts | [h-c.ai](https://www.h-c.ai/) — what is verified, what is in progress, and what is not available | | Machine-readable | [facts.json](https://www.h-c.ai/facts.json) and [llms.txt](https://www.h-c.ai/llms.txt), generated from the same ledger as the pages | | Measurements | [research](https://www.h-c.ai/research) and [benchmarks](https://www.h-c.ai/benchmarks), with raw output retained | | Limits and guarantees | [limitations](https://www.h-c.ai/limitations), [security](https://www.h-c.ai/security) and [reliability](https://www.h-c.ai/reliability) | | Using it | [docs](https://www.h-c.ai/docs) — the interface as it stands, kept in step with the build | As of 2 September 2026, h-c.ai lists nine items as verified, each with raw measurement output retained. They are reproduced here by title only. The wording of each, and anything added since, is at h-c.ai, and a script in this site’s build compares this list with the live facts.json before every deployment. - A machine-readable retrieval contract - Evidence at paragraph level, verbatim - Citations survive re-indexing - Tamper-evident chunk text - One gate for everything - Search evidence is a verbatim slice of the original - Evidence points back to the original file - Shared access with keys and grants - One gate, exercised on Windows and Linux - Spreadsheets become documents ## The work this suits, and the work it does not What decides whether a knowledge base like this belongs in a given line of work is not its size or its speed. It is that an answer arrives with the passage it came from, unchanged, and that the first stage is deterministic: the same question returns the same passage for as long as the corpus and the version are unchanged. Everywhere that property is worth paying for, it is worth paying for because somebody eventually asks why the system said what it said. **Regulated and audited work** is the clearest case. In legal, accounting, insurance and pharmaceutical practice, an answer that cannot be traced to the document it came from is not evidence of anything; it is a suggestion. The auditor’s question is never “is this plausible” but “show me the clause”. Public-sector work has the same shape for a different reason: a decision affecting a citizen has to be explicable to that citizen. **Work that is not allowed to leave the building** is a category rather than an industry, and it is larger than it looks. Defence suppliers, hospitals, corporate legal departments, and overseas subsidiaries on connections that drop for an afternoon all share one requirement: the knowledge stays where it is. The local form runs as one file on one machine with no network socket open; the shared form runs on a machine you choose, and nothing is hosted by us. **Japanese and other CJK document estates** are a fit for a narrower, more technical reason. Indexing by character trigrams removes the morphological analyzer from the query path, and with it a dependency that has to be maintained, tuned per domain, and blamed whenever a search misses. What it does not remove is the need to turn a question into a query: search takes keywords, and a sentence is not a query. Retrieval quality on Japanese has no published benchmark of its own yet, so a pilot on your own documents is the only honest way to find out. > Where this is the wrong tool. If your users ask questions in words that appear nowhere in the documents, the half of stage two you need is the half that is not connected. If you need a sentence typed by a person to work as a query with nothing in between, that belongs to the calling agent today and not to the database. If you need a hosted service, we do not offer one; shared access means running one process yourself. In each of those cases the honest answer is a different design, and we would rather say so here than discover it with you in month three. ## The same argument we make to clients, applied to ourselves Everything on the [ontology page](https://www.tech-japan.jp/services/ontology/) says that retrieval fails upstream of the model, and that the fix is structure: controlled terminology, defined relations, and a path you can trace. Hadano AI Cabinet is that argument compiled into software — a retrieval engine where the traceable path is the architecture rather than a feature bolted on afterwards. It is also why the accounting on this page stops where it does. A knowledge tool that quietly returns a plausible empty answer is worse than one that errors, and a product page that quietly keeps last month’s specification is the same failure in a different place.