EvidenceAgent-MM v0.1.0
EvidenceAgent-MM is an evidence-grounded multimodal system for noisy meetings and classrooms. It answers questions such as who proposed what, when, why, and which slide was visible with claim-level citations, timestamps, anonymous speaker/page provenance, confidence decomposition, and a tool trace. When evidence is ambiguous or insufficient, the system asks a targeted clarification or abstains.
This repository is a system model card and reproducibility artifact, not a newly trained neural checkpoint. EvidenceAgent-MM v0.1.0 composes pinned upstream models behind a deterministic evidence gate. No upstream weights are copied into this repository and no fine-tuning is claimed.
- Source: https://github.com/Jatshi/EvidenceAgent-MM
- Release: https://github.com/Jatshi/EvidenceAgent-MM/releases/tag/v0.1.0
- Chinese guide: https://github.com/Jatshi/EvidenceAgent-MM/blob/main/README.zh-CN.md
- From-scratch Chinese tutorial: https://github.com/Jatshi/EvidenceAgent-MM/blob/main/docs/tutorials/evidenceagent_mm_from_scratch_tutorial.md
System composition
| Role | Upstream component | Exact revision / model | Verified use |
|---|---|---|---|
| Generation | Qwen/Qwen3-8B |
b968826d9c46dd6066d109eabc6255188de91218 |
evidence-constrained wording after gating |
| Dense retrieval | BAAI/bge-m3 |
5617a9f61b028005a4858fdac845db406aefb181 |
multilingual evidence ranking |
| ASR | Systran/faster-whisper-small |
536b0662742c02347bc0e980a01041f333bce120 |
timestamped transcript atoms |
| OCR | PaddleOCR 3.7 | PP-OCRv5_mobile_det + PP-OCRv5_mobile_rec |
screen/slide evidence atoms |
| Core gate | EvidenceAgent-MM | v0.1.0 | Answer / Clarify / Abstain control flow |
Consult each upstream repository for its own model license and usage restrictions. EvidenceAgent-MM code and the system-level configuration in this repository are Apache-2.0.
What the “model” is
The core contract is not “send a video to an LLM.” It is:
- ASR, OCR, and turn detection create typed
EvidenceAtomobjects. - SQLite FTS5, deterministic/dense embeddings, reciprocal-rank fusion, and a bounded graph retrieve evidence.
- A deterministic gate checks ambiguity, required modalities, retrieval quality, and claim support.
- Only evidence that passes the gate reaches the optional Qwen renderer.
- The response is one of
answered,needs_clarification, orabstained.
The exact thresholds and pinned components are machine-readable in system_config.json.
Verified results
All measurements below use one CC0 synthetic 12.4-second meeting unless explicitly marked otherwise. They validate integration and contracts, not real-meeting accuracy.
| Check | Result |
|---|---|
| EAMM-Bench Bronze | 12 sessions / 120 questions |
| Three-state status accuracy | 1.000 |
| Evidence Recall@5 | 1.000 |
| ECE-10 | 0.413, uncalibrated |
| Core tests | 29 passed, 88.08% branch-aware coverage |
| faster-whisper small | 2 segments, WER 0.125, 1.587 s warm-cache |
| BGE-M3 | correct target at rank 1, score 0.625, 7.710 s |
| Qwen3-8B | both citations and required facts preserved, 12.201 s total |
| Qwen3-8B peak VRAM | 15,665 MiB |
| PaddleOCR | 6 atoms from 2 slides, 6 unique IDs, 2.461 s |
| Turn-detection fallback | 2/2 turns, temporal IoU 0.914; not speaker identity |
| AutoDL deterministic API | 200 requests, 0 failures, 234.5 req/s, P95 137.0 ms |
Raw JSON reports are stored in results/. Exact environment freezes and SHA-256 provenance are stored in provenance/.
Quick start
git clone https://github.com/Jatshi/EvidenceAgent-MM.git
cd EvidenceAgent-MM
python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dev]'
make benchmark
eamm --db /tmp/eamm.db serve --host 127.0.0.1 --port 8000
The web demo has no authentication. Keep it on localhost unless you add an authenticated reverse proxy and a retention policy for meeting media.
To download this Hub artifact itself:
from huggingface_hub import snapshot_download
snapshot_download(repo_id="jatshi/EvidenceAgent-MM", local_dir="EvidenceAgent-MM-hub")
The pinned upstream models can be downloaded independently with their exact revisions. This preserves attribution, lets the Hub deduplicate official files, and avoids presenting third-party checkpoints as EvidenceAgent-MM training output.
Known limitations
- Bronze is a deliberately easy synthetic contract benchmark and is not evidence of production accuracy.
- Confidence is a hand-built baseline score; ECE-10 shows that it is not calibrated.
- The synthetic ASR output confuses
reviewwithrenewand includes a number-format difference. - Mobile OCR misses
42 msand readslatencyasIatencyon the first slide. - The license-free fallback detects speech turns only. Its sequential labels are not persistent speaker identities.
- Citation presence is not the same as semantic entailment; production use needs claim-level support evaluation on real data.
- Private meeting media requires access control, consent, deletion policies, and an audit log.
Citation
@software{evidenceagent_mm_2026,
author = {Shi, Jianting},
title = {EvidenceAgent-MM: Evidence-Grounded Multimodal Assistance for Noisy Meetings},
year = {2026},
version = {0.1.0},
url = {https://github.com/Jatshi/EvidenceAgent-MM}
}