The interface
The app has three sidebar destinations: Chat, Files, and Reports. Here, window workspace means how Chat and the optional Notebook are arranged in the window. The title-bar switcher chooses a mounted folder (called a workspace in saved records); that folder is the data boundary and does not determine the window layout.
Notebook stays in the same app window. On a narrow window, the panes adapt to the available width; opening Notebook does not change source scope or analytical state.
The durable work
The normal path is lightweight: ask in Chat, then open Notebook when the work benefits from a brief, a Question plan, source choices, run history, or saved Findings. The Investigation can be named, edited, resumed, and reviewed independently of the transcript.
Workspace context and knowledge
Choose Workspace context from the title-bar workspace actions to editfella.md and review knowledge built from analysis. The guide contains user-authored definitions and instructions. Analysis-derived facts retain their supporting run and evidence; a proposal can stay within its originating Investigation, be edited, be shared across the workspace, or be dismissed. Investigation definitions remain scoped to their Investigation unless the user explicitly promotes a fact.
Notebook organization
An Investigation’s main sections are:- Brief — its title, objective, and working context.
- Sources — what the workspace contains, source readiness, and the files chosen for the work.
- Questions — the root question, its subquestions, and their answer states.
- Analysis — saved runs, methods, source revisions, evidence, and checks.
- Results — the current answers, saved Findings, caveats, and coverage.
Analytical changes
An analytical diff compares a saved output revision with another revision of that same output. It answers five questions:- Meaning: Did the question, definition, filter, or assumption change?
- Inputs: Did the source selection or source revision change?
- Method: Did the saved computation or presentation change?
- Results: What values changed, when the two result sets can be safely aligned?
- Impact: Which explicit report placements or checkpoints refer to the earlier revision?
Generated answers
When a question has useful structured results, Fella asks the model to compose a responsive JSON-render presentation using registered components such as prose, metrics, tables, charts, callouts, and presentation-only controls. The model chooses a layout suited to the question and results. It does not get to invent a new data source or execute code in the renderer: analytical values bind to saved result IDs and fields, and the client validates the presentation before rendering it. Plain-language answers remain valid when a structured view is unnecessary or cannot be composed. Reopening a saved answer uses its saved presentation and results rather than rerunning the analysis.Retention and deletion
Deleting a conversation removes its transcript and linked Investigation. Saved artifacts, reports, their revisions, and the analysis records needed to support saved work remain in that workspace. The confirmation dialog explains this before deletion. Removing a workspace from the sidebar also leaves its folder and files untouched; reconnect the folder to open its saved reports and results.Current boundary
The app currently uses one selected provider model for a conversation. Fella does not provide automatic model routing or a Fella-hosted inference service. The Rust engine remains responsible for data access, computation, evidence, and verification; the model directs the investigation and composes the explanation and presentation.Investigations
Create and review a persistent analytical investigation.
Reports and saved outputs
Preserve results, compare revisions, and export a report.
Models and providers
Connect a provider and choose a model for the conversation.
Architecture
See how the model, Rust engine, persistence, and renderer fit together.