Category guide

What is a Knowledge Compiler?

The compile contract

The unit of value is not a chat response. It is a Compiled World with immutable inputs, structured objects, exact evidence, review state, retrieval material and portable files.

One Compiled World, from the published sampleOpen the full World

12 of 6,310 objects drawn · 1,281 evidence regionssha256 aff67d5c6d0…

Compiled from Apple's public SEC filings, so every result can be re-derived. One issuer's clean, English-language filings at a single permission level show the mechanism; they are not a claim about a mixed internal corpus or a degraded scan. What this read recovers, and what it does not, is in the capability manifest.

  1. INPUTVersioned source files and connector cursors.IMMUTABLE
  2. STRUCTUREDirectory, ontology, entities, claims and relations.REVIEWABLE
  3. EVIDENCESource version, the exact location inside it, excerpt and digest.TRACEABLE
  4. WORLDCandidate and active revisions separated by a human decision.VERSIONED

Where each category acts

These are not four competing products. They are four different spans of one pipeline, which is why the comparisons below are about scope rather than quality.

Where each category acts on the compile pipelineOne pipeline in six stages: sources, read, structure, evidence, world, projections. Enterprise search acts at sources, finding the document. A knowledge graph is the structure stage, storing objects and relations. RAG acts at projections, retrieving chunks at question time. A Knowledge Compiler spans all six: it compiles, binds evidence to regions, and versions the result.THE COMPILE PIPELINESOURCESREADSTRUCTUREEVIDENCEWORLDPROJECTIONSEnterprise searchfinds the documentKnowledge graphstores objects and relationsRAGretrieves chunks at question timeKnowledge Compilercompiles, binds evidence to regions, and versions the result
  1. SOURCESEnterprise search finds the document
  2. READ
  3. STRUCTUREKnowledge graph stores objects and relations
  4. EVIDENCE
  5. WORLD
  6. PROJECTIONSRAG retrieves chunks at question time
  7. ALL SIXKnowledge Compiler compiles, binds evidence to regions, and versions the result

When a compiler is the right tool

Compiling costs a reading pass and a review. It pays for itself where the same material is answered from repeatedly, where being wrong is expensive, and where somebody will eventually ask how an answer was reached.

  1. REPEATED USEThe same corpus answers many questions, for many people, over months.PAYS BACK
  2. CONSEQUENCEA wrong answer costs a return visit, a rework order, a filing or a claim.WORTH PROVING
  3. SCRUTINYSomebody will ask which page an answer came from, and a plausible sentence is not an acceptable reply.NEEDS EVIDENCE
  4. CHANGESources are revised, superseded and amended, and last quarter's answer still has to be explainable.NEEDS VERSIONS

When it is not the right tool

Hide when it is not the right tool

Cases this product does not serve.

  1. ONE DOCUMENTA single file you will read once. Open it. The compile buys nothing you do not already have.USE A READER
  2. LIVE RECORDSAnswers that are a query over a database or a ticket queue, not a claim written in a document.QUERY THE SYSTEM
  3. NO REVIEWERNobody who can decide whether a candidate World is correct. Activation is a human decision by design, and without one the World never becomes active.NEEDS A PERSON
  4. BEYOND THE SOURCESQuestions whose answer is not in the material. Retrieval here is a matching test, not a judgement about whether what matched answers you: it declines only when nothing matched at all, so a question the corpus cannot answer comes back as the nearest matching regions with their locators rather than as a refusal. Judging whether the matched regions answer your question stays with the reader.OUT OF SCOPE

Glossary

Open glossary

The words this product uses in the exact sense it uses them. Where a term has a looser industry meaning, the narrower one here is deliberate.

  1. COMPILED WORLDThe output of one compile: objects, relations, evidence, retrieval material and a validation report, addressed by a digest.THE ARTIFACT
  2. CANDIDATEA compiled result that has not been activated. It can be read, downloaded and reviewed; nothing answers from it.LIFECYCLE
  3. ACTIVEThe one revision a workspace answers from. It changes only when a person activates a candidate or rolls back to a prior revision.LIFECYCLE
  4. EVIDENCE REGIONA source version, an exact location inside it, an excerpt and a digest. What a citation resolves to. The location takes whatever form the source has; in a PDF it is a page and a box measured in thousandths of it.PROVENANCE
  5. STABLE IDAn identifier derived from content, so recompiling the same source lands on the same object rather than a new one.IDENTITY
  6. MANIFEST DIGESTThe sha256 of the canonical form of a World. Two Worlds with the same digest are the same World.IDENTITY
  7. ABSTENTIONThe answer given when the sources do not support one. It is a result, not a failure.BEHAVIOUR
  8. PACKAGEThe portable form: canonical model, Turtle, JSON-LD, CSV, retrieval JSONL, provenance and a validation report.PORTABILITY

Questions people ask

Hide questions people ask

Short answers about what the product does today.

Is this just RAG with extra steps?
RAG is a retrieval strategy; this is an artifact. The chunks a retriever needs are one file in the package, produced from a reviewed World rather than from raw text — so they carry the page and region they came from, and they change only when the World does.
Do I have to replace my retrieval stack?
No. The package ships the graph as Turtle, JSON-LD and CSV, and the retrieval units as JSONL, so an existing vector store, graph database or agent framework loads them without adopting anything else.
What happens when a source document changes?
The new bytes are a new version, and compiling produces a new candidate rather than editing the World in place. The active revision moves only when a person activates it, and the previous revision stays readable so an older answer remains explainable.
What does it do when it does not know?
It declines when nothing in the World matched the question at all, and says which sources it looked at. What it does not do is decide whether what matched actually answers you — that judgement stays with the reader, which is why every answer carries the page and the region it came from. A composed answer with no region behind it would be indistinguishable from a correct one, and that is the failure this whole contract exists to prevent.
What stops the output from being a black box?
Every object carries the regions that support it, the package is a set of open formats, and the validator that checks a package is a readable script rather than a service — so a package can be verified without asking us anything.

The package is the contract

Portability is only real if someone outside can check it. The package format, its required files and the two verifiers — one for the archive's signature, one for what is inside it — are documented, and a compiled sample is open without an account.