ChatBHT Research Notes

ChatBHT Research Notes#

  • Local ModelsInstitutional infrastructureNeeds: Computation · Energy
  • SAIA APIGWDG · KISSKINeeds: Shared infrastructure
  • External ModelsCommercial providersNeeds: API costs
University
AI ApplicationFrontend options Data · Access · Budgets
Students · Educators · Researchers
Knowledge Potential

Explanations · Graphics · Solutions · Code

Resources investedCompute · Energy · API fees · Human time

“A broad access to capable AI leads to decentralized innovation.”

Students already use commercial AI for coursework. An institutional service provides equitable access while keeping responsibility for data, usage rules and budgets with the university.

We need a data economy for AI-generated knowledge.

When someone develops a good explanation, creates a useful graphic with AI, or finds an interesting solution, work has gone into it. Computing resources were used, time was invested, and human decisions were made.

Yet much of this knowledge disappears into personal chat histories.

Knowledge Cycle#

Prompt, Process, Provenance

Research Data Alliance · framework proposal

ChatBHT

  1. Prompt

    Define the question and context.

  2. Process

    Generate, evaluate and revise.

  3. Provenance

    Preserve the record for future work.

How did the result come about?

  • Which prompt was used?
  • What context was provided?
  • What output came back?
  • Which parts were adopted?
  • What did the human change afterwards?

Design Principles#

AccessEqual access

Low-barrier access to capable models, accessible interfaces, responsible-use teaching and transparent cohort budgets.

AutonomyRoom to explore

Permissionless innovation within agreed data and resource boundaries. Members can address institutional problems from within.

TrustTransparent use

Transparent model choices, clear data handling and deliberate sharing. Private by default; no hidden prompt analysis or automated student rankings.

ArchiveLasting knowledge

Preserve reviewed results with their context, version and permissions. Make them searchable and reusable without losing their origin.

Knowledge Architecture#

  1. 01

    Versioned Archive

    Git keeps a history of saved changes; GitLab makes these files accessible as projects with access permissions. ChatBHT already archives conversations, metadata and agent files there. A commit link points to one exact saved version, so later changes do not obscure what was originally recorded.

  2. 02

    Automatic Updates

    GitLab System Hooks are automatic notifications from the archive server. When a project is created or new files are uploaded, an enabled hook can notify a processing service. The proposed service would check permissions and queue approved changes across projects, instead of checking every archive manually.

  3. 03

    Structured Records

    An extractor would read changed archives and turn the recorded details into a consistent entry: prompt, model, context, output and human contribution, where available. Each entry would retain its source version. Results from different projects could then be searched and compared in the same way.

    Proposed record fields
    • interaction_id
    • timestamp
    • subject
    • model
    • model_version
    • prompt
    • context_sources
    • output
    • tool_calls
    • artefacts
    • token_usage
    • visibility
    • license
    • human_review
    • human_changes
    • commit_id
  4. 04

    Searchable Catalogue

    A separate search store would catalogue approved entries, avoiding the need to reopen thousands of projects for each question. It would support keyword searches, topic similarity and aggregate statistics. Changes to permissions or deleted source material would also have to be reflected in the catalogue.

  5. 05

    Linked Knowledge

    A knowledge graph would connect entries through named relationships: which source was used, which dialogue produced a graphic, or where a result was reused. Similarity could suggest related work; human review would distinguish these suggestions from verified connections.

    • PromptresemblesPrompt
    • Artefactderived fromInteraction
    • InteractionusesSource
    • Resultreused inProject
  6. 06

    Institutional Search

    An AI-assisted search would answer questions using the institution's approved archives and link back to the recorded versions. It could reveal related work across departments and partner institutions. Personal introductions would be opt-in; access permissions would still apply to every result.

    Which MQTT security approaches have already been developed at our university?

    Which prompts produced useful explanatory graphics for mechatronics?