File size: 3,775 Bytes
eff511c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 | ---
license: mit
title: MitrAnlil Pilot3
sdk: gradio
emoji: 🏢
colorFrom: yellow
colorTo: yellow
short_description: YOK Kanunlar için normatif AI asistanı
sdk_version: 6.20.0
---
# MitrAnlil MCKF v0.8 Multi-Document Corpus
MitrAnlil, yükseköğretim mevzuatı için kaynak gösterimli normatif bilgi asistanıdır. Bu paket v0.7 pre-release mimarisini koruyarak 2547 tek belge yapısından üç belgeli MCKF corpus yapısına geçer.
## Indexed Documents
- `TR-KANUN-2547` — 2547 sayılı Yükseköğretim Kanunu
- `TR-KANUN-2914` — 2914 sayılı Yükseköğretim Personel Kanunu
- `TR-KANUN-2809` — 2809 sayılı Yükseköğretim Kurumları Teşkilatı Kanunu
Belge registry:
```text
data/registry/legal_documents.json
```
Runtime tek corpus kaynağını yükler:
```text
data/mckf/corpus_mckf_ontology.json
```
2547 per-document çıktıları geriye dönük audit ve regresyon için korunur; legacy chunk/semantic-unit/canonical-concept dosyaları runtime’a geri sokulmamıştır.
## Build
```bash
python tools/build_mckf_from_source.py --all
python tools/build_mckf_from_source.py --document TR-KANUN-2914
python tools/build_mckf_from_source.py --document TR-KANUN-2809
```
No-arg legacy mode yalnız `TR-KANUN-2547` çıktısını yeniden üretir:
```bash
python tools/build_mckf_from_source.py
```
## Runtime Flow
```text
conversation_intent
-> source_router
-> corpus evidence retrieval with document/article scoring
-> deterministic fallback or LLM synthesis from selected evidence only
-> full article/source rendering
-> feedback logging
```
The source router scores candidate documents before retrieval. Returned candidates expose scoring details including source-router score, document-domain score, semantic-role score, norm-type score, lexical score, structure score, evidence-focus score, cross-document bonus, wrong-document penalty, and generic-article penalty.
## RDFox Pilot Layer
Bu sürümde RDFox mimarisi MVP'ye harici RDFox servis bağımlılığı olmadan indirgenmiştir. `corpus_mckf_ontology.json` tek runtime kaynağı olarak kalır; RDFox uyumlu pilot katman bu kaynaktan üretilir:
```text
data/mckf/rdfox_pilot_manifest.json
data/mckf/rdfox_pilot_triples.nt
data/mckf/rdfox_pilot_rules.dlog
```
Pilot katman şu dar çekirdeği kapsar:
- Document ingestion
- Semantic Normative Unit parser
- MCKF builder
- Hybrid retrieval: semantic-role/dense-ready, BM25/lexical, concept-role
- Evidence sufficiency gate
- Source-locked generation
- Citation renderer
Runtime'da `rdfox_pilot.py` beş micro-rule sağlar: deadline priority, procedure chain, exception warning, cross-reference follow ve definition attach. Bu kurallar retrieval skor kırılımına `pilot_rule_score` ve `pilot_rule_hits` olarak yansır; RDFox/Datalog karşılığı `rdfox_pilot_rules.dlog` içinde tutulur.
## Verification
Run from this package root:
```bash
python -m compileall .
python tools/build_mckf_from_source.py --all
python tools/audit_mckf.py --all
python tools/run_golden_checks.py
python tools/run_multidoc_golden_checks.py
python tools/runtime_health_check.py
python tools/export_feedback_regression_candidates.py
```
Local verification uses `..\.venv\Scripts\python.exe` when `python` is not on PATH.
## Known Limits
- Local verification does not execute external Groq synthesis because `GROQ_API_KEY` is not present locally.
- 2914 schedule sections are detected as separate `ScheduleUnit` source units and are retrievable; row-level table semantics are not yet normalized into typed cells.
- The semantic role extractor remains heuristic.
- The system is a source-grounded explanation assistant, not legal advice.
|