Assessment: L1 pipeline for L2 and Post-Processing

[vprashrex]

Is your feature request related to a problem?
Currently, all rows are sent directly to the L2 LLM batch without any pre-filtering. This leads to wasted L2 costs on irrelevant or duplicate submissions, and the output lacks flexibility in terms of computed columns and sorting.

Describe the solution you'd like

• Implement an L1 pipeline for topic relevance and duplicate detection before processing with L2 using a Celery task.
• Ensure topic relevance evaluates selected attachment documents, with relevance metrics included in export.
• Enable detection of per-item attachment types (supporting mixed image and PDF columns).
• Apply post-processing features (computed columns, filter, sort) at the export stage.

Original issue

Describe the current behavior
All rows go straight to the L2 LLM batch. No pre-filter, so irrelevant/duplicate submissions consume full L2 cost. Relevance ignores attachments, columns can't mix image+PDF, and export output is fixed (no computed columns/filter/sort).

Describe the enhancement you'd like

• L1 pipeline (topic relevance + duplicate detection) running before L2 via a Celery task.
• Topic relevance that also evaluates selected attachment documents, with per-input relevance in the export.
• Per-item attachment type detection (mixed image/PDF columns).
• Post-processing (computed columns, filter, sort) applied at export.

Why is this enhancement needed?
Cuts L2 cost by dropping irrelevant/duplicate rows early, improves relevance accuracy by judging real document content, and gives reviewers derived/filtered/sorted exports.