Week 3:- Metadata Extraction Chain

[Abhishek-Kumar-Rai5]

Objective

Implement the metadata extraction pipeline that converts normalized Document objects into populated Intermediate Representation (IR) instances using LLM-based structured extraction.

Extraction follows BETYdb's dependency hierarchy and progressively builds the IR.

---

Planned Work

Metadata Extraction

Implement sequential extraction for:

1. Citation
2. Site
3. Treatment
4. Species

Each stage will consume only the document blocks relevant to its task.

Prompt Design

Develop structured prompts for:

• citation extraction
• site extraction
• treatment reconstruction
• species identification

Treatment extraction will reconstruct experimental design rather than perform simple entity extraction.

LLM Integration

Integrate the Jetstream inference endpoint through the project's inference abstraction.

Initial target models:

• DeepSeek R1
• Llama 4 Scout

Provenance

Every extracted field must retain:

• supporting evidence
• originating document objects
• confidence
• reasoning metadata where available

Evaluation

Evaluate extraction on the five ground-truth papers.

Produce:

• preliminary accuracy metrics
• failure taxonomy
• documented extraction limitations

---

Deliverables

• Working metadata extraction chain
• IR population
• Prompt templates
• Evaluation on all ground-truth papers
• Initial accuracy report