QiEvidence Domain (QiEn)
Purpose
QiEvidence (internal engine name: QiEn) is a local-first evidence intelligence module that transforms raw multimedia and documents (videos, photos, PDFs, audio, screenshots, etc.) into a structured, cross-referenced, searchable evidence archive.
Core capabilities: - Timestamped transcription and visual event extraction - Neutral, cold, objective narration (no emotion, no speculation, no mind-reading) - Entity normalization and tracking (people, vehicles, locations, accounts, etc.) - Automated contradiction / inconsistency detection - Master timeline assembly across all artifacts - Human review queue for uncertain or conflicting items
It is a disciplined evidence-processing engine with two brains: - Extractor: Factual, structured output only. - Analyst: Cross-referencing, flagging, and neutral narration.
Folder Structure
The local execution environment uses a standardized folder hierarchy:
/QiEvidence/
├── input/
│ ├── videos/
│ ├── photos/
│ ├── documents/
│ └── other/
├── working/
│ ├── audio/
│ ├── frames/
│ └── transcripts/
├── output/
│ ├── events/
│ ├── entities/
│ ├── timelines/
│ ├── reports/
│ └── reviews/
└── archive/ (hashed originals + metadata)
Schema: qievidence
The domain uses the following tables, linked to qione.tenants and qicase.cases.
| Table | Description |
|---|---|
evidence_records |
Master artifact registry (one row per file) |
evidence_segments |
Chunked slices of long media/documents |
evidence_entities |
Canonical people, vehicles, locations, accounts, etc. |
evidence_entity_mentions |
Join table (where each entity appears) |
evidence_events |
Observed or extracted events |
evidence_claims |
Statements/assertions extracted from content |
evidence_observations |
Machine-grounded visual/textual facts (non-claims) |
evidence_links |
Cross-references between any two objects |
evidence_conflicts |
Detected inconsistencies (the “money” table) |
evidence_timelines |
Named timeline containers |
evidence_timeline_items |
Ordered items in timelines |
evidence_reviews |
Human review queue |
Processing Pipeline
- Ingest: Hash file, read metadata, assign evidence ID, store in
evidence_records. - Extract:
- Audio →
faster-whisper→ transcript. - Video →
ffmpegkeyframes →Qwen2.5-VLobservations. - Photos/Docs → OCR +
Qwen2.5-VLdescription. - Structure: Populate segments, entities, events, claims, observations.
- Link & Compare: Build links and run contradiction scanner.
- Timeline Assembly: Populate timelines and items.
- Narrate: Final cold, logical report using strict prompt rules.
- Review: Push uncertain/conflicting items to
evidence_reviews.
Narration Rules
- No emotion words.
- No speculation or intent inference.
- Separate Observed / Stated / Uncertain.
- Preserve timestamps and source IDs.
- Mark unclear audio/visuals explicitly.