创新 · 数据与分析
从 11,000 页卷宗中抽出 10,000 个实体的图谱。
创新部门持有自由文本卷宗，分析几乎用不上。生成式抽取 POC 产出了可探索、可审计的结构化图谱。
关键指标
- 客户
- Bpifrance
- 行业
- 创新
- 职能
- 数据与分析
- 状态
- POC 已交付 — 工业化进行中
- 部署
- 托管云（客户租户）
背景
Hundreds of dossiers, more than 11,000 pages: qualitative natural-language data largely unused for cohort analysis or public-policy indicators.
问题
- Links between startups, labs, patents and funding were unstructured.
- Manual analyses were long and non-reproducible. No continuously usable overview.
方案
Co-designed ontology (node and relation types), generative extraction, relational storage, visualisation, expert correction loop. The Evidence Panel cites source sentences back to the original document.
调用的 NEXA 组件
| 组件 | 作用 |
|---|---|
| Arbitration algorithm | Orchestration of extraction models |
| Evidence Panel | Source sentences back to the document |
| HITL | Validation / correction of nodes and relations |
| AI Knowledge Vault | Co-designed, reusable ontology |
| Cockpit | Precision, coverage, footprint |
影响
| 维度 | 具体影响 |
|---|---|
| Data | Quantitative ecosystem indicators inaccessible before structuring. |
| Precision | 80% precision; no hallucinations observed in control. |
| Scalability | Architecture extensible to other corpora. |