QuEra Computing builds neutral-atom quantum computers. The company came out of Harvard and MIT research, and its machines are open to the public through Amazon Braket. Its physicists, engineers, and leadership need to know what the company has already established, and they cannot afford to be told something it never did.
From April 2025 to September 2026 we built and ran a private knowledge graph for them over the company's material, deployed inside its AWS account. It answers with citations or declines to answer at all.
The answer exists. Getting back to it is the problem.
A research organisation has the information. It lacks a way to get back to the specific thing somebody already established. "What did we conclude about this approach, and what was it based on" has an answer. It lives in a paper, an internal write-up, an experiment log, a Slack thread, or one person's memory of a meeting.
The default move in 2025 was to point a language model at the pile and ask. For a research team that fails in the worst way. The model answers fluently whether or not the evidence exists. A confident paraphrase of a result nobody measured is worse than no answer, because someone will act on it before they discover it was wrong.
Every answer traceable to the document that supports it. A system that declines when the support is not there.
the requirement, as we wrote it down
Every place the company writes things down, in one graph
The knowledge a researcher wants back is spread across tools with different shapes. Wikis and papers have hierarchy: pages, headings, tables, figures. Chat and meetings have time: who said what, after what. We treat the two differently on the way in, so a citation can point at a heading in one case and at a timestamp in the other. Either way, the reader can open the source and check.
- NotionPages, blocks, and nested databases.parse · block · chunk · graph
- ConfluenceEngineering wiki pages with headings and tables.parse · block · chunk · graph
- Uploaded documentsPDFs and Office files: papers, specs, lab records.parse · block · chunk · graph
- SlackTime-ordered messages across channels.index · chunk · search
- ZoomMeeting recordings as time-coded transcripts.index · chunk · search
Ingestion runs continuously. A research corpus changes every week, and a graph that falls behind it is a graph people stop trusting.
It cites, or it declines
We built a knowledge graph with a gate in front of the answer. Documents are parsed into entities and the relationships between them, and every relationship carries two clocks: when the fact was true in the world, and when the system learned it. A physicist can ask what the company believed as of a date. A lead can follow a claim back to where it started and forward to everything built on it since.
- 01Ingest
A connector pulls new and changed content from each source on a schedule.
- 02Parse
Documents become blocks with a path: page, heading, paragraph, table cell.
- 03Chunk and embed
Blocks become chunks with character offsets back into the block they came from.
- 04Extract
Entities and relationships are pulled from each chunk, with the chunk recorded as evidence.
- 05Link in time
Each relationship gets a business-time and a system-time interval. New evidence re-scores its confidence.
- 06Retrieve
Six signals are combined and ranked, then the reader resolves each hit back to its full source.
- 07Answer, or decline
The assistant writes with citations attached, or reports that the evidence is not there.
People do not ask questions the way a search box wants them to. A researcher types an abbreviation, a colleague's first name, and a half-remembered phrase from a meeting, and expects all three to land. So retrieval combines six ways of matching instead of betting on one.
- exactCanonical name match. "AWS" finds "AWS".
- keywordBM25 full-text search over every chunk.
- chunkEmbedding similarity over raw chunk content.
- entityEmbedding similarity over the entities in the graph.
- link
Relationships are embedded too, so "who owns X" matches the edge, not a hop.
- alias
Abbreviations, nicknames, and an imported company glossary.
The assistant remembers between sessions. Decisions and facts from earlier conversations are captured in the background, the current task and open questions are recorded explicitly, and a small set of skills runs on the same graph: a weekly digest, meeting preparation, a knowledge-gap report, an onboarding guide. When the question is about numbers, the answer arrives as a chart, a metric, or a table rather than a paragraph.
> Which approach did we settle on for the pipeline rewrite, and what was it based on?
The team settled on the staged approach in November 2025. The decision
followed a review of two options and was recorded after the design
meeting on the 4th; the Slack thread two days later confirms the owner.
[1] Confluence · Pipeline rewrite: options review · 2025-11-04 · §Decision
[2] Slack · #eng-platform · 2025-11-06 · thread, 3 messages
[3] Notion · Q4 engineering plan · 2025-11-10 · block /page/2/paragraph/4
> What did the vendor benchmark show for option C?
// No evidence in the connected sources mentions a benchmark for option C.
// Closest material: [1] above, which lists option C as not evaluated.A citation you can open, not a link you have to trust
The part of the engagement that took longest was making every citation hold up when somebody clicks it.
Most systems keep two separate pictures of a company's knowledge. One is the documents: the page, the heading, the paragraph. The other is the graph: the entity, the relationship, the confidence score. The document store can show you the text and the graph can tell you what relates to what, and nothing ties a relationship back to the exact line that justified it. The citation is decoration on top of a guess.
For a research organisation that gap is the whole problem. "The team concluded X" is worth something only if you can open the paragraph where they concluded it, see the date, and see what has been written since that agrees or disagrees.
That is what makes the citation in Exhibit 05 more than a link. It is a path from relationship to evidence to chunk to block to page to source. When a physicist challenges an answer, the system shows its work.
Nothing leaves the account
An unpublished result, a partnership under NDA, and an internal assessment of a competitor's approach are the material a quantum computing company most needs to search and can least afford to send to somebody else's inference endpoint. So the system runs inside QuEra's own AWS account, and it was designed that way from the start rather than hardened into it later. Models and index sit where the data already lives. No question, document, or embedding crosses the account boundary to produce an answer.
- EKSKubernetes cluster running the API, the workers, and the ingestion pipeline.
- Bedrock · ClaudeManaged model access for answers, inside the account.
- KubeAISelf-hosted embeddings and local models on the cluster's own GPUs.
- SurrealDBGraph, chunks, evidence, and bi-temporal links in one store.
- S3Source files and parsed artifacts.
- KongGateway and authentication in front of the API.
Outside the frame: nothing in the answer path. No third-party inference endpoint, no external index, no shared tenancy.
For a board or a security review, the answer path has no component outside the account, adds no vendor, and shares no tenancy. When the client's team asks what the system can reach, the diagram above answers it.
What QuEra says
He is very professional, knowledgeable and great to work with.
Hugh Kelly, Fractional CMO at QuEra Computing · LinkedIn, September 2026
What this page does not claim
This case study quotes no metric for the effect of the system. We have internal figures for corpus size and retrieval performance, but we measured them ourselves and have not confirmed them with the client. A number a buyer cannot check is not evidence.
The figures in Exhibit 01 describe the engagement. When QuEra puts its name to a result, it goes here. Until then, this page claims only what we can defend: what we built, why we built it that way, and where it runs.
If your organisation holds knowledge it cannot send outside its own walls, this is the shape of the system we would build for you. Book a call and bring the question your team keeps failing to get an answer to.