Having extensive material and using it well are different capabilities. A lasting AI employee may face old descriptions, project rules, experience, and current messages. A larger context window does not decide which content belongs in this work.
At the start, the product needs to help prepare a useful basis for the current question: the goal, confirmed conditions, required facts, and missing evidence to find. A user can then keep asking questions without assembling all the background into another long message each time.
Teloa organizes project-background preparation and material retrieval separately, then records what actually enters the turn. Preparation carries forward established working conditions; retrieval finds further evidence for the immediate question. Together, they reduce material transfer and provide a path back when an answer goes wrong.
Prepare known background, then investigate the specific question
Project goals, existing briefs, active rules, and required material should not depend on being rediscovered by each search. Work linked to a project prepares these first. Permitted optional material and memory are then selected against the current question, scope, and input budget. Current project selection uses local word and Chinese-character-fragment matching; it is distinct from semantic retrieval in the library.
Retrieval addresses a different need: knowing what to check without knowing which passage holds the evidence. When preparing a pilot quotation, project background can establish that only the pilot is under discussion. The next question may concern seat limits in the service material. Permitted semantic retrieval can return relevant excerpts and fixed locations so the employee can check whether they apply to this quotation.
The division connects what the work is meant to achieve with where its evidence can be found. Background alone may produce a correctly directed answer without evidence. Search alone may accurately explain an annual contract while overlooking the pilot-only scope. Preparing the direction and then checking specific conditions provides a better basis for usable analysis.
Define the permitted scope, then compare relevance
Suppose two job descriptions are being checked against hiring requirements. The descriptions, checking method, and project rules may all help. A semantically similar employment contract from another business may rank highly while remaining outside the material allowed for this task.
Retrieval therefore identifies candidates. Sources, ownership, versions, and scope continue to constrain their use. Relevance as the only criterion can make search seem intelligent while bringing inappropriate context into an answer.
Local semantic retrieval currently requires a selected model and prepared material. Results contain fixed content locations, limited excerpts, and coverage, identifying what was searched and what is not yet ready.
Coverage directly affects the conclusion. If only two prepared job descriptions were searched, the analysis should concern those two. A third that is not yet searchable needs to remain listed as unchecked.
Relevance ranking asks which passage may help. Material governance asks whether this work may obtain it. Both conditions must hold in that order. Supplying unrelated client material and merely telling the model not to use it has already brought unsuitable content into the decision process.
Selected material still needs a record of which turn received it
The project mechanism prepares goals, rules, and selected material, but preparation is not sending. Actual dispatch records show which content entered which model request, keeping an unexecuted preparation from being counted as use.
This distinction is particularly useful in continuing work. Material added in the second turn cannot be retrospectively treated as the basis for the first turn’s conclusion. If the material changes later, the older answer still needs the version available at the time.
The skills currently installed and the files currently in the knowledge library cannot establish what earlier work actually used. Configuration describes what is available now. Usage records explain a piece of work that has already happened. They need to be examined separately.
Citations make checking easier
These records answer different questions:
| Record | What it establishes |
|---|---|
| Dispatched input | The material entered that request |
| A retrieved excerpt | The query returned that candidate |
| A completed tool call with a receipt | The operation produced the corresponding result |
None of these records alone proves that the model used the material sentence by sentence, or interpreted it correctly. A source may say “for pilots only” while the output describes a general effect. Further retrieval tuning may not fix that problem of understanding and expression.
Citations shorten the path back to the evidence. The user can inspect the fixed source and its conditions, then decide whether to correct the input, change the method, or redo part of the conclusion. A name that looks like a link but cannot lead back to the actual evidence does not serve that purpose.
Prepare later inputs again when sources change
When a source changes, future input needs reassessment. Earlier turns retain their original basis; later reads check the current state. Project source constraints also cover summaries and derived content, allowing their later use to end when the original becomes invalid.
This adds relationship and state maintenance, but gives corrections a chance to affect later work. Updating material or withdrawing experience should not require ever-longer prompts to counter hidden copies.
Context can be evaluated in sequence: was the material appropriate, did it reach this turn, and did the output interpret it correctly? Capacity, relevance, and final quality each matter; one number cannot answer for the other stages.