Skip to content

Fella v0.2.0

Choose a tag to compare

View all tags
@Avijit-Kumar-GIT Avijit-Kumar-GIT released this 20 Sep 19:10
· 475 commits to main since this release
v0.2.0
a6c5922

Fella v0.2.0 is the focused personal analytics release. It keeps Fella
local-first, read-only, and BYOK while making the path from a user's files to
an answer more capable and easier to inspect.

Highlights

  • Ask-first local analytics workspace with sources, context, conversations,
    settings, light/dark appearance, keyboard navigation, and provider setup.
  • Query-derived bar and line charts with longer time-series support, exact
    values, source evidence, and verification state.
  • Embedded WASM RustPython analytics through Wasmi with bounded execution, no
    filesystem/network/environment/subprocess access, and a read-only SQL bridge.
  • Experimental capability policy switches for table, document, Python, and
    visualization analysis paths.
  • Inspectable evidence, provenance, workspace revisions, query replay, and
    deterministic verification checks.
  • Documentation hosted at https://docs.lilfella.app.

Improvements

  • The analytics engine now supports longer time-series charts without applying
    the old small-category limit, while keeping category charts readable and
    rejecting truncated or incomplete visual data.
  • Small query results can be shown completely to the model, with bounded
    previews and explicit truncation when a result is too large.
  • Multi-file context, shared join columns, written-month date normalization,
    atomic workspace snapshots, and model-loop cancellation make analysis more
    dependable across real folders.
  • The model loop keeps step limits, bounded tool output, concurrent independent
    tool calls, and a final deterministic verification pass.
  • Provider setup is clearer for hosted BYOK connections, including the model
    identity shown in the Ask footer.

Bug fixes

  • Named-month dates such as Aug 1, 2026 are normalized consistently in CSV,
    JSON, and Excel inputs instead of collapsing date groupings into a blank
    bucket. Ambiguous numeric dates remain unguessed.
  • Answer self-checks now flag wrong aggregate operations and date/time groupings
    that collapse to a null bucket.
  • Case-sensitive filters on likely category or label columns receive guidance
    so values such as Leisure and leisure are not silently mismatched.
  • Learned SQL recipes are no longer reused as durable memory, and overlapping
    user corrections supersede earlier notes instead of creating contradictions.
  • Oversized Python output is labeled as truncated, and user cancellation of a
    running Python analysis remains responsive.
  • Chart evidence, source provenance, and verification status stay attached to
    generated visual answers rather than being lost in rendering.

Product boundary

The plugin, packs, augment, and marketplace runtime are removed from the
default product surface. /mcp remains an inert experimental command and no
official MCP connector ships in this release.

Release notes

This is an unsigned early release. The embedded WASM/RustPython boundary is an
application guardrail for generated analysis code; the longer-term OS-level
sandbox design remains documented for future work. Manual packaged-build smoke
testing and the frozen hosted-model question battery remain follow-ups.

Assets 9
Loading