structsense/tutorial/python-example/ner-config.yaml at main · sensein/structsense · GitHub

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
# agent config
agent_config:
  extractor_agent:
    role: >
      Neuroscience NER Extractor Agent
    goal: >
      Extract named entities and key terms from neuroscience text {input_text}.  Return structured JSON.
    backstory: >
      You are an AI assistant for neuroscience NER. You do not hallucinate. Output strict JSON: entities and key_terms.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

  alignment_agent:
    role: >
      Neuroscience NER Concept Alignment Agent
    goal: >
      Add mapping info to {extracted_structured_information}: map each entity/resource/key term to an ontology. Add ontology_id, ontology_label, ontology, and concept_mapping_provenance ("tool" if from Concept Mapping Tool, "llm_knowledge" if from your knowledge).
    backstory: >
      You align extracted terms to ontologies (e.g. CL, UBERON, NCBITaxon). You have the Concept Mapping Tool; use it. Always set concept_mapping_provenance to indicate source (tool vs llm_knowledge). Output strict JSON.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

  judge_agent:
    role: >
      Neuroscience NER Judge Agent
    goal: >
      Extend {aligned_structured_information} by checking quality and adding judge_score (0–1) and remarks for each item.
    backstory: >
      You evaluate alignment quality. Add judge_score and remarks; do not remove existing fields. Output strict JSON.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

  humanfeedback_agent:
    role: >
      Neuroscience NER Human Feedback Agent
    goal: >
      Take human feedback and apply it to {judged_structured_information_with_human_feedback}. Revise the result using the human's input.
    backstory: >
      You incorporate human feedback into the judged result. Preserve structure; apply corrections and notes from the human. Output strict JSON.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

# task config (pipeline variables: input_text → extracted_structured_information → aligned_structured_information → judged_structured_information_with_human_feedback)
task_config:
  extraction_task:
    description: >
      Extract entities and key_terms from the input text. Use the NER tool on {input_text}. Return JSON with entities (entity, label, sentence, start, end, paper_location, paper_title, doi) and key_terms.
    expected_output: >
      JSON: { "entities": [...], "key_terms": [...] }
    agent_id: extractor_agent

  alignment_task:
    description: >
      Add mapping info to the extraction. Input: {extracted_structured_information}. Use the Concept Mapping Tool to map each entity/resource/key term to an ontology. Add ontology_id, ontology_label, ontology, and concept_mapping_provenance to each mapped concept. Set concept_mapping_provenance to "tool" when the mapping comes from the Concept Mapping Tool (BioPortal), or "llm_knowledge" when you provide the mapping from your own knowledge without calling the tool. Preserve all existing fields; only add mapping info and provenance.
    expected_output: >
      Same structure as input, with ontology_id, ontology_label, ontology, and concept_mapping_provenance ("tool" | "llm_knowledge") per entity/item.
    agent_id: alignment_agent

  judge_task:
    description: >
      Extend the aligned result by checking quality. Input: {aligned_structured_information}. Add judge_score (float 0–1) and remarks for each item. Preserve all existing fields (entity, label, ontology_id, ontology_label, sentence, start, end, etc.); add judge_score and extend remarks.
    expected_output: >
      Same structure as input, with judge_score and remarks added per item.
    agent_id: judge_agent

  humanfeedback_task:
    description: >
      Take human feedback and apply it. Input: {judged_structured_information_with_human_feedback}. The human may provide corrections or notes; incorporate them into the result. Preserve structure and add/update remarks as needed.
    expected_output: >
      Same structure as input, revised according to human feedback.
    agent_id: humanfeedback_agent

# embedding config
embedder_config:
  provider: ollama
  config:
    api_base: http://localhost:11434
    model: nomic-embed-text