Keep bindings and handlers
Web components and desktop markup retain attributes, scripts, styles, and events. The reader does not render the interface or resolve framework bindings.
Evidence and limits
Follow the claim.
Read its source.
An answer should give you enough evidence to check it. OmniTome keeps source identity and citations attached to retrieved knowledge, and offers optional checks for generated claims and question coverage.
Source coverage
A declaration is evidence, not execution.
Software readers retain exact authored spans across UI, application, interface, database, and infrastructure sources. The parse census records which reader ran and whether analysis was partial or used fallback.
Keep authored contracts
RPC schemas, SQL definitions, and recognized Entity Framework mappings remain source evidence. They do not establish the live service contract or deployed database schema.
Keep static references
Bicep, ARM, Terraform, Kubernetes, and Helm retain supported declarations and authored references. No reader deploys, renders templates, or resolves live resources.
From result to answer
Checks with different jobs.
Deterministic retrieval and model-generated answers are separate paths. The model checks below are opt-in; they add model calls and latency.
Keep the original source
Each citation carries its locator, revision, scope, and content hash. A composed atom is included only when the sealed audience is entitled to every required input.
Assess the selected excerpts
The optional support check distinguishes an explicit negative answer from an undocumented capability. Code verifies that a selected supporting span exists in the served source.
Assess claims together
The optional composition check binds each proposed claim to its chosen excerpts, checks question coverage, and allows bounded repairs. A refused draft stays out of the final answer.
Evaluate your own knowledge base
Ask the questions that matter to your work.
Source quality, question phrasing, retrieval settings, and model choice affect usefulness. Include answerable questions, comparisons, and questions your sources cannot answer.
Same input, same output, bit for bit?
Compare retrieval under fixed files, settings, and embedding identity. Generated answers and external model calls have separate variability.
Did the right documents make the five-result short list?
Write down which sources answer a question before searching. Check whether they appear, as well as which relevant sources were missed.
Could you finish the job from what is on the list?
Inspect prerequisites, steps, exceptions, and each side of a comparison. A related passage may still leave out the part needed to act.
How much of what it showed you was worth your time?
Count useful results separately from merely related material. Broader retrieval can improve coverage while adding irrelevant results.
Does it know when it does not know?
Include unsupported capabilities, wrong scopes, and incomplete evidence. An explicit negative statement can answer a question; missing documentation cannot establish that negative.
Is the first good result near the top?
Record the position of useful evidence. Exact vector search means exact distance computation; it does not guarantee semantic relevance.
Where did that claim come from?
Read each claim against its selected quotes. Check that the source describes the same operation, subject, and conditions as the answer.