August 16, 2026. I want VALESKA to return a useful section, not hand me a whole document and call the job finished. VALESKA is the document and organizational-memory system I am building. Tonight's repository changes add section-based knowledge units, a way to store them, and a rights-tagged source selection for internal work. That is a concrete engineering milestone. It is not yet evidence that an assistant answers better.
If you have just inherited your company's AI initiative, this distinction is worth keeping. An embedding is a numerical representation used to find related text. It does not, by itself, establish where an instruction begins, which warning belongs beside it, or whether the material may be reused. Finding similar words and preserving usable evidence are different jobs.
What I mean by a unit
A knowledge unit is a bounded piece of source material that retains enough identity and context to be examined separately. For a manual, that can be a heading and its section body, with its place in the document hierarchy and a connection back to the source. It is not a fact certified as true. A neatly bounded section can still be wrong, outdated, or inappropriate for a particular reader.
Today's section handling follows document headings, including numbered manual sections, rather than treating an arbitrary length of text as the natural boundary. The committed tests cover keeping a section together, preserving parent context, and not confusing a heading-like line inside a code example with a real heading. Here, manual means a document such as a handbook; it does not mean I have hand-labelled every section.
The storage code keeps the unit's full text, but the text sent for embedding can be shortened to fit the model. That limit matters. Preserving the whole section in storage does not guarantee that every sentence influences how search finds it. A long procedure with its decisive warning near the end is a useful difficult case, not something I can wave away because the unit has a title.
A label is not permission
Rights tags are metadata recording intended uses and restrictions. Today's source-selection code marks material for internal use and against public release or transfer to third parties. Those labels are not legal licence proof, an ownership investigation, or evidence that consent has been obtained. A company still needs a defensible basis for using the underlying documents. Calling a collection internal does not settle that question.
The export code selects source units and filters navigation material, short entries, and repeated text. It does not generate training examples, call a language model, or train one. The source slice is an internal research artifact, not a public dataset. OpenText manuals are explicitly outside this slice until they have real section units. I am working with my documentation here, not reporting a customer corpus converted and approved.
Corpus provenance means the recorded origin and processing history of the collection: where material came from, how it was selected or divided, and which source a returned unit belongs to. Today's units carry source and processing context to support that trace. Provenance helps someone inspect a claim; it does not make the claim correct or substitute for rights review.
The next test
The evidence tonight is committed implementation and test cases, not a measured retrieval improvement or a verified training run. My next test is deliberately small: ask a question whose answer sits in a known section, retrieve it, and check the section against the original. Then repeat with a long section whose important detail is near the end. I also want to inspect the restrictions beside the selected source text and check what was excluded. If I cannot explain those results to the person responsible for the company's documents, adding more documents is not progress.
Historical basis: VALESKA commits 2e7ced0, e11a9ef, and 43a1f0e, August 16, 2026, America/Los_Angeles. These establish the section, storage, and source-selection implementations; no live execution is claimed here.