resource-extraction-config.yaml

# BBQS Resource Extraction - agent config

agent_config:
  extractor_agent:
    role: >
      Brain and Behavior Quantification and Synchronization (BBQS) Resource Extractor Agent
    goal: >
      Extract a single structured resource entry per input relevant to BBQS. Capture related subjects as secondary mentions.
    backstory: >
      You are a research assistant for the BBQS consortium (NIH-funded: tools, datasets, models, benchmarks for brain-behavior research).
      Extract the primary resource only; reference other entities (datasets, benchmarks, tools) under a `mentions` field. Do not hallucinate. Return valid JSON only.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

  alignment_agent:
    role: >
      BBQS Concept Alignment Agent
    goal: >
      Align extracted resource metadata with BBQS controlled vocabularies and domain schemas.
    backstory: >
      You normalize resource fields (type, category, target, specific_target) using existing schemas and ontologies (e.g. BTO, NCBITaxon). Correct typos and ambiguities. Return one updated resource in JSON; mentions may be left as-is or annotated.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

  judge_agent:
    role: >
      BBQS Judge Agent
    goal: >
      Evaluate {aligned_structured_information} against BBQS standards and output structured JSON with a confidence score and rationale.
    backstory: >
      You are a reviewer for the BBQS resource pipeline and do not hallucinate. Score only the primary resource (not mentions). Return confidence 0–1, rationale, and preserve original fields with evaluation metadata in JSON.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

  humanfeedback_agent:
    role: >
      BBQS Human Feedback Processor Agent
    goal: >
      Evaluate {judged_structured_information_with_human_feedback} and incorporate human expert feedback.
    backstory: >
      You act as a human domain expert reviewer for BBQS and do not hallucinate. Score only the primary resource. Return confidence 0–1, rationale, and preserve fields; process user feedback. Output JSON.
    llm:
      model: openrouter/openai/gpt-4o-mini
      base_url: https://openrouter.ai/api/v1

# agent task config
task_config:
  extraction_task:
    description: >
      Extract structured metadata for one BBQS resource from {input_text}. One resource per input; put other referenced resources under `mentions` only.

      Return JSON with: name, description, type (Model, Dataset, Paper, Tool, Benchmark, Leaderboard), category, target, specific_target, url, mentions (optional: models, datasets, benchmarks, papers). Use null if missing. Do not copy the example verbatim.

      Resource:
      { input_text }

    expected_output: >
      output format: json
      Example:
      "extracted_resources": {
        "1": [{
          "name": "DeepLabCut Model Zoo - SuperAnimal-Quadruped",
          "description": "Pre-trained model for quadruped animal pose estimation using DeepLabCut.",
          "type": "Model",
          "category": "Pose Estimation",
          "target": "Animal",
          "specific_target": "Quadruped, Horse, Mice",
          "url": "https://deeplabcut.github.io/DeepLabCut/docs/ModelZoo.html",
          "mentions": {
            "datasets": ["Quadruped-80K", "AnimalPose"],
            "benchmarks": ["AP-10K", "AnimalPose"],
            "models": ["DeepLabCut"]
          }
        }]
      }

    agent_id: extractor_agent

  alignment_task:
    description: >
      Align the extractor output to BBQS concepts. Map entities to matching concepts from existing resources/ontologies (e.g. BTO, NCBITaxon). If no match, use N/A.

      Inputs: {extracted_structured_information}

      Add mapped_target_concept (id, label, ontology) for target and mapped_specific_target_concept per specific_target. Preserve name, description, type, category, target, specific_target, url, mentions.

    expected_output: >
      output format: json
      Example:
      "aligned_resources": {
        "1": [{
          "name": "DeepLabCut Model Zoo - SuperAnimal-Quadruped",
          "description": "Pre-trained model for quadruped animal pose estimation.",
          "type": "Model",
          "category": "Pose Estimation",
          "target": "Animal",
          "mapped_target_concept": [{"id": "http://purl.obolibrary.org/obo/BTO_0000042", "label": "animal", "ontology": "BTO"}],
          "specific_target": "Quadruped, Horse, Mice",
          "mapped_specific_target_concept": [
            {"specific_target": "Mice", "mapped_target_concept": {"label": "Mus musculus", "id": "NCBITaxon:10090", "ontology": "NCBITaxon"}}
          ],
          "url": "https://deeplabcut.github.io/DeepLabCut/docs/ModelZoo.html",
          "mentions": {"datasets": ["Quadruped-80K"], "benchmarks": ["AP-10K"], "models": ["DeepLabCut"]}
        }]
      }

    agent_id: alignment_agent

  judge_task:
    description: >
      Evaluate {aligned_structured_information}: (1) Assess alignment quality with BBQS standards, (2) Assign judge_score 0–1, (3) Append judge_score to the resource. Preserve all aligned fields.

    expected_output: >
      output format: json
      Example:
      "judge_resource": {
        "1": [{
          "name": "DeepLabCut Model Zoo - SuperAnimal-Quadruped",
          "description": "Pre-trained model for quadruped animal pose estimation.",
          "type": "Model",
          "category": "Pose Estimation",
          "target": "Animal",
          "mapped_target_concept": [{"id": "http://purl.obolibrary.org/obo/BTO_0000042", "label": "animal", "ontology": "BTO"}],
          "specific_target": "Quadruped, Horse, Mice",
          "mapped_specific_target_concept": [{"specific_target": "Mice", "mapped_target_concept": {"label": "Mus musculus", "id": "NCBITaxon:10090", "ontology": "NCBITaxon"}}],
          "url": "https://deeplabcut.github.io/DeepLabCut/docs/ModelZoo.html",
          "mentions": {"datasets": ["Quadruped-80K"], "models": ["DeepLabCut"]},
          "judge_score": 0.95
        }]
      }

    agent_id: judge_agent

  humanfeedback_task:
    description: >
      Evaluate {judged_structured_information_with_human_feedback} as human expert: (1) Assess alignment with BBQS standards, (2) Assign judge_score 0–1, (3) Apply user feedback. Do not reduce the number of extracted resources.

      modification_context: {modification_context}
      user_feedback_text: {user_feedback_text}

    expected_output: >
      output format: json
      Same structure as judge_task output (judge_resource with "1", "2", ...). Preserve all fields; add or update judge_score and process user_feedback_text in remarks/rationale if needed.

    agent_id: humanfeedback_agent

# embedding config
# https://docs.crewai.com/concepts/memory#additional-embedding-providers
embedder_config:
  provider: ollama
  config:
    api_base: http://localhost:11434
    model: nomic-embed-text