One HMRC booklet, three readers, and each one sees only what their role is cleared to see.
Ask, don't search
Traditional search hands you a list of candidates and leaves the reading to you. That's fine when there are only a few documents to flip through, but expensive when the answer is buried across six contracts, two policy updates, and three years of internal notes and you've no idea which one to open first.
Orrery reads every document for you and gives you the answer in plain language, with the exact source attached: not a best match, not a summary, just the answer and the line that settles it. And when your own documents disagree, it shows you the contradiction instead of quietly picking one.
Built different
Orrery's AI works from the facts in your documents and how they connect, not the keywords a search engine matches, and not the word-by-word prediction behind a chatbot. Ask a question and it returns the answer with the exact source it came from. Nothing generated, nothing inferred, nothing invented.
That distinction runs deep: you can show your working, not just hand over an answer; contradictions between documents are surfaced rather than smoothed over; the same question always returns the same answer; and access is enforced where the data sits, not by an instruction someone could talk a model out of.
hallucinations
Answers trace to a source
guesswork
It looks up, it never guesses
black boxes
Every step is on the record
Traceability
Corroboration & contradictions
Cost per query
Throughput
Natural language quality
Knowledge separation policy
Bring your own data
Real-time ingestion
Hallucination risk
Deterministic results
Data privacy
Audit trail
| Orrery | LLM | Elasticsearch | |
|---|---|---|---|
| Traceability | Every answer resolves to a named source you can cite and re-check | None no link between the wording and any specific source | Passage-level highlights the exact passage it matched |
| Corroboration & contradictions | Structural agreeing sources reinforce each other; conflicting ones are flagged at answer time | Unreliable compares passages in context, but misses conflicts at scale and states one version confidently | No retrieval-only; no comparison between documents |
| Cost per query | Negligible runs on ordinary servers, not specialised AI hardware | High GPU mandatory · $2–8 per 1k queries at scale | Low server cost only · no GPU needed for keyword |
| Throughput | 400+ q/s sustained on a single ordinary server | 2–5 q/s A100 GPU · serial token generation per request | 200–500 q/s keyword search is fast; vector adds a little latency |
| Natural language quality | Structured answers precise and cited, not prose | Fluent prose best for open-ended consumer Q&A | Document snippets returns ranked passages; reader interprets |
| Knowledge separation policy | Structural enforced in the data itself; no query can override it | Prompt-only soft constraint · jailbreakable · not auditable | Infrastructure-level real access rules, enforced by the system, not a prompt |
| Bring your own data | Shared facts two large corpora at a fraction of the combined cost, not double | Context window hard cap ~1M tokens · cost grows linearly | Additive index re-index on schema change · can't relate records to each other |
| Real-time ingestion | Live new facts queryable the moment they're added | Training cutoff new data requires fine-tuning or RAG scaffolding | Near-live ~1s refresh · segment merge adds some tail latency |
| Hallucination risk | None answers resolve to stored facts; no text generation | Inherent model fabrication is structural, not a configuration bug | None retrieval-only · returns indexed content verbatim |
| Deterministic results | Always same query + same data → same answer | Not guaranteed temp=0 helps but GPU float variance across batches still drifts | Always deterministic ranking · reproducible across requests |
| Data privacy | Sealed on request cloud-hosted · never used to train a model · sealed hardware on request | Cloud-first data processed by provider · fine-tuning may use your inputs | Self-hostable fully on-premise · open source · no vendor data access |
| Audit trail | Per-query log who asked · which sources matched · what scope was applied | Black box token attribution impossible without RAG scaffolding | Slow-log only query-level timing · no semantic attribution |
LLM figures based on a hosted frontier model at standard API pricing. Elasticsearch figures based on a standard vector-search deployment. Orrery figures measured on a single commodity server under sustained load.
License
Commercial enquiries
Free lets you try Orrery on a small set of documents. Commercial covers business use, with integrations and email support. Enterprise adds the heavy machinery: fine-tuning to your terms, the latest models, a hardware enclave that not even our staff can see into, and the kind of dedicated support and SLAs your auditors expect.
Get in touch and we'll talk through licensing for your organisation.
Contact us →Free
Commercial
Enterprise