Repository navigation
FollowThrough v0.3.0 — semantic evidence and visual conflict review
FollowThrough v0.3.0 — semantic evidence, visual sources and cited conflict review
Download followthrough-v0.3.0.zip, extract, create a Python 3.11+ virtual environment, install requirements.txt and run run.py. Open http://127.0.0.1:8787. The frontend is already built in the ZIP.
Load guided demo for a clearly labelled fictional walkthrough without accounts or credits. Live features require your own Nebius Token Factory key and YAML model configuration. No developer key or private user data is distributed.
New in this release:
- Qwen embeddings shortlist approved, unexpired evidence; NVIDIA Nano reviews relevance. Candidates remain suggestions until selected and reviewed. The documented dedicated Qwen reranker returned model-not-found, so reranking explicitly uses NVIDIA chat relevance assessment.
- NVIDIA Super VL transcribes scanned PDF pages, PNG/JPEG and document tables. Compare the transcription against each original page and confirm or correct it before analysis. At most six pages per visual import. This is model transcription with human review, not guaranteed OCR accuracy.
- On-demand NVIDIA Super reasoning produces source-cited conflict and amendment findings. Findings require owner decisions; stale reviews block approval/export. No raw chain-of-thought is displayed or stored.
Core features retained: sourced requirements, PDF/DOCX/text imports, approved company evidence, missing-input tasks, amendment review, approval invalidation, DOCX/attachment export, separate submission and receipt evidence.
Validation: 11 standalone automated tests, including real HTTP/SQLite contracts, plus three live synthetic semantic/vision/reasoning smoke cases and six earlier extraction cases. These are not a production accuracy benchmark or customer impact measurement. See submission/advanced-live-evaluation.json and submission/live-evaluation.json.
Original core-workflow YouTube demo (v0.2): https://www.youtube.com/watch?v=TGrqRwF3qm0
This video does not show the v0.3 additions. New feature screenshots are included in submission/.
Single-owner local application. No public multi-tenant hosting, legal guarantee, automated external submission, arbitrary spreadsheet ingestion or independently authenticated receipts. Semantic evidence and visual transcriptions require human review. Gmail OAuth is optional and unverified. Do not expose this local app through a public tunnel.
See README.md for setup and model settings, submission/SUBMISSION.txt for the proposed Devpost entry, and docs/ARCHITECTURE.md for boundaries.