LLApp is a coaching-support application used by school leadership staff — principals, instructional coaches, and administrators — to plan coaching conversations and generate related documents. This card describes the AI system that powers LLApp's chat and planning features: what it is, what it is intended for, and what its known limitations are. LLApp does not train its own model; it uses a hosted, third-party large language model to generate responses.
LLApp is an application built around a hosted large language model, not a model LLApp trains or owns. It uses:
| Function | Model used | Purpose |
|---|---|---|
| Conversation and document generation | A general-purpose large language model | Drives the coaching-assistant chat and drafts planning documents |
| Request classification | The same general-purpose model | Determines what supporting information a request needs before responding |
| Image text extraction | A vision-capable version of the same model | Reads text from images embedded in uploaded materials, including handwriting |
| Semantic search | A text-embedding model | Powers similarity search over the knowledge base used for retrieval |
Not applicable. LLApp is delivered only as a hosted application; there is no public repository, downloadable model, or demo associated with this card.
As implemented, the AI system supports:
Output generated by the system (coaching scripts, feedback documents, lesson evaluations, calendars) is intended to be exported and used outside the application — for example, printed or shared with staff. The system does not track how exported content is subsequently used.
The system is intended for coaching-meeting planning, instructional-data review, culture and routine design, observation feedback, and related staff communications. It has not been evaluated for, and is not intended for, use in making decisions about individual students or staff (for example, discipline, evaluation ratings, or personnel actions) without independent human judgment.
Human review is recommended before acting on AI-generated content in any situation with real consequences for a student or staff member.
Not applicable. LLApp is accessed only as a hosted application; there is no standalone model or package for external integration.
Not applicable. LLApp does not train or fine-tune a model. It maintains an internal knowledge base of reference materials that is indexed for retrieval and given to the model as context at answer time — this material informs individual responses but does not train or modify the model itself.
Not applicable, for the same reason.
Not applicable to model training. Reference materials added to the knowledge base are broken into smaller sections and indexed for retrieval before being made searchable.
Not applicable.
Not applicable.
No held-out benchmark dataset is used to evaluate the quality of the AI system's responses.
Not defined. No evaluation broken out by subject area, grade level, or user population has been conducted.
No formal accuracy, faithfulness, or hallucination-rate metrics are currently tracked.
Not applicable — no benchmark results exist for this system.
There is no automated or benchmark-based evaluation of response quality at this time. Instead, users can rate individual responses and flag concerns (for example, inaccurate or off-topic answers, overly generic responses, or missed follow-up actions); this feedback is reviewed and selectively used to improve future responses. This is an ongoing quality-improvement process, not a formal evaluation.
Not applicable. The underlying model is accessed only through the provider's API; no interpretability analysis of the model itself has been conducted.
Not measured. LLApp does not operate the infrastructure used to run the underlying model — inference is performed by the AI provider, which does not publish per-request energy or carbon figures that this card can report.
At a high level, LLApp responds to a user request in three stages: first, it determines what supporting information (if any) the request needs; second, it gathers that information from relevant sources — prior conversation, the knowledge base, or the user's own uploaded files; and third, it provides the request and gathered information to the underlying language model, which generates the response using a defined set of allowed actions (such as creating a document or looking up a reference guide). Long conversations are periodically condensed into a shorter summary so that the conversation stays within the model's working context.
Not disclosed in this card.
Not applicable.
Not applicable.
Compiled from internal review, dated 2026-08-12.
Not specified in this card. Contact your organization's LLApp representative for further information.