Work · Knowledge & Context

Make knowledge usable.

The knowledge already exists. The challenge is getting the right part into the work with enough history, authority and context to trust it.

Bring us the problem

← All work

Where knowledge lives.

A company's knowledge is spread across documents, tickets, repositories, dashboards, chat threads and the people who remember why something was decided. Very little of it was written to be read by a system.

Storing more of it does not fix this. A well-organised document store still cannot tell you which of two contradicting policies governs the case in front of you, whether the answer it found is still current, or whether the person asking is allowed to see it.

That gap is where most knowledge projects quietly stall: the search works, and the work still needs a human to check it.

Retrieval is not context.

RAG and search are genuinely useful, and we build them. Given a question they find passages that look related, which is a hard problem and a solved-enough one.

But finding and applying are different operations. Retrieval answers "what looks relevant?" Context also has to answer "what applies here?" — and that question needs things a similarity score does not carry.

Which source has authority when two disagree. Whether this version is current. What has already happened in this case. What the person or agent is permitted to see. And what should be deliberately left out.

The context pipeline.

The architecture depends on the work, but the shape of it is usually this. Each stage is a decision, not a service you can buy.

Sources Documents, tickets, code, systems, and the state of the case in flight.

Retrieval and selection What could be relevant, narrowed to what the task can actually use.

Authority, permissions, state Which source governs, whether it is current, and what this user may see.

Task context The assembled set that is worth sending — and what was deliberately excluded.

Model or agent The step everyone focuses on, and the one that depends most on the four above.

Trace What was used, where it came from, and why the answer can be checked.

What stays in people.

Some context will not be captured, and it is worth being honest about which. Why an exception was granted three years ago. Which document everyone knows is out of date. What a customer is actually worried about underneath the ticket they filed.

We are not trying to ingest an organisation. We are trying to preserve much more of the decisions, history and provenance the company already produces, and to make what it has trustworthy at the moment of work.

Selection is efficiency.

Better selection is not a cost optimisation bolted on at the end. It is the same act as better relevance, seen from the resource side: material that never should have been sent also never has to be paid for, waited on, or filtered out by the model.

Fewer irrelevant tokens means lower latency and cost, and usually a better answer, because a model given a smaller and more governed set has less to be wrong about. Where selection, reuse or measurement of context is part of the problem, LeanCTX can provide that layer.

What we can build.

Finding

RAG and hybrid retrieval
Metadata and structure
Ranking against the task, not the query

Governing

Source authority and freshness
Permissions and access
Provenance and traceable answers

Remembering

Case and task state
Memory across sessions
Context selection and reuse

Measure the result.

The metric comes from the task rather than from the technology, and it is chosen before the system changes.

It may be answer quality, time to an answer someone will act on, how much manual verification remains, whether a claim can be traced to a source, or the volume of context required to get there.

Quality · Time to trust · Verification · Traceability · Tokens

The one worth watching most is verification: a system that is right and still gets double-checked has not changed the work.

What should it know?

Tell us what the system is trying to do and where the knowledge lives today.

Start a conversation

hello@thinkery.ch