Create agents.py

agents.py

@@ -0,0 +1,192 @@
+# agents.py
+import os, json, textwrap, pathlib
+from typing import Dict, Any, Tuple
+from openai import OpenAI
+
+from files_process import load_input_text
+
+# ---------- Instructions ----------
+
+def build_case_study_instructions() -> str:
+    return textwrap.dedent("""\
+    You are a senior medical writer specializing in AI in healthcare.
+    Using ONLY the provided INPUT (no fabrication), produce a professional case study in **Markdown**.
+
+    STRICT REQUIREMENTS (cite inline in prose, not as links):
+    - Use established reporting guidance:
+      • CARE case reports for completeness/transparency (CARE, 2013/2017).
+      • CONSORT-AI (trial reports) and SPIRIT-AI (protocols) where applicable (2020).
+      • TRIPOD+AI (2024) for prediction model reporting (discrimination, calibration, validation).
+      • HIPAA de-identification Safe Harbor: state that 18 identifiers are removed or avoided; avoid re-identification risk.
+      • FDA AI/ML SaMD perspective: risk controls, monitoring, change management.
+    - If a section lacks data, write “Not specified.”
+    - Use a neutral, clinical tone (avoid marketing fluff).
+    - Prefer short paragraphs and tables where appropriate.
+
+    MANDATORY SECTIONS (use these exact headings):
+    # Title
+    ## Executive Summary
+    - 3–6 bullets
+
+    ## Clinical Context & Problem
+    ## Patient/Population & Setting
+    ## Data Sources & Governance
+    - Provenance; access; quality checks; de-identification approach (HIPAA Safe Harbor)
+    - Security/compliance: HIPAA (and GDPR if applicable)
+
+    ## AI/ML Approach
+    - Task definition; target(s)
+    - Features/data preparation
+    - Model(s); training/validation split; external validation if any
+    - Fairness/bias checks
+
+    ## Evaluation & Metrics if provided
+    - Classification/regression metrics
+    - Calibration; confidence estimation
+    - Clinical outcomes (if available)
+    - Reference TRIPOD+AI for what to report
+
+    ## Workflow Integration & Safety
+    - Human oversight; failure modes; alerting
+    - Monitoring & model updates (FDA SaMD AI/ML perspective)
+
+    ## Results & Impact
+    - Clinical impact; operational efficiency; ROI/costs where applicable
+
+    ## Ethics & Bias Mitigation
+    ## Regulatory, Privacy & Security
+    - HIPAA/GDPR; access controls; audit
+    ## Limitations & Generalizability
+    ## Conclusion
+
+    LENGTH: aim 4,000–5,800 words.
+
+    OUTPUT: Valid Markdown only.
+    DO NOT include code fences around the Markdown.
+    """)
+
+def build_manager_instructions() -> str:
+    return textwrap.dedent("""\
+    You are the manager reviewing three case study drafts on AI in healthcare.
+
+    TASKS:
+    1) Rate each draft on a 1–10 scale for:
+       - Clinical completeness (CARE)
+       - AI reporting rigor (TRIPOD+AI)
+       - Trial/protocol framing where relevant (CONSORT-AI / SPIRIT-AI)
+       - Privacy & regulatory correctness (HIPAA Safe Harbor; FDA SaMD)
+       - Clarity & structure
+    2) Briefly justify each rating (1–3 sentences).
+    3) Pick a **single winner** among the three drafts (best overall).
+
+    OUTPUT JSON (strict):
+    {
+      "scores": [
+        {"agent": "agent1", "clinical_completeness": int, "ai_rigor": int, "trial_framing": int, "privacy_regulatory": int, "clarity_structure": int, "justification": "..."},
+        {"agent": "agent2", "clinical_completeness": int, "ai_rigor": int, "trial_framing": int, "privacy_regulatory": int, "clarity_structure": int, "justification": "..."},
+        {"agent": "agent3", "clinical_completeness": int, "ai_rigor": int, "trial_framing": int, "privacy_regulatory": int, "clarity_structure": int, "justification": "..."}
+      ],
+      "winner": "agent1|agent2|agent3"
+    }
+
+    IMPORTANT:
+    - Return only JSON.
+    """)
+
+# ---------- Agent calls (OpenAI SDK) ----------
+
+def _openai_client() -> OpenAI:
+    return OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))
+
+def call_openai_case_study(input_text: str, model: str = "gpt-4o-mini") -> str:
+    client = _openai_client()
+    system = build_case_study_instructions()
+    prompt = f"INPUT:\n{input_text}\n\nFollow the instructions strictly."
+    resp = client.responses.create(
+        model=model,
+        instructions=system,
+        input=prompt,
+        temperature=0.3,
+    )
+    return resp.output_text.strip()
+
+def call_gemini_case_study(input_text: str, model: str = "gpt-4.1-nano") -> str:
+    client = _openai_client()
+    system = build_case_study_instructions()
+    prompt = f"INPUT:\n{input_text}\n\nFollow the instructions strictly."
+    resp = client.responses.create(
+        model=model,
+        instructions=system,
+        input=prompt,
+        temperature=0.3,
+    )
+    return resp.output_text.strip()
+
+def call_deepseek_case_study(input_text: str, model: str = "gpt-4.1-mini") -> str:
+    client = _openai_client()
+    system = build_case_study_instructions()
+    prompt = f"INPUT:\n{input_text}\n\nFollow the instructions strictly."
+    resp = client.responses.create(
+        model=model,
+        instructions=system,
+        input=prompt,
+        temperature=0.3,
+    )
+    return resp.output_text.strip()
+
+def call_openai_manager(agent1: str, agent2: str, agent3: str,
+                        model: str = "gpt-4o") -> Dict[str, Any]:
+    client = _openai_client()
+    manager_instr = build_manager_instructions()
+    payload = {
+        "agent1_draft": agent1,
+        "agent2_draft": agent2,
+        "agent3_draft": agent3,
+    }
+    resp = client.responses.create(
+        model=model,
+        instructions=manager_instr,
+        input=json.dumps(payload),
+        temperature=0.2,
+        response_format={"type": "json_object"},
+    )
+    raw = resp.output_text
+    try:
+        return json.loads(raw)
+    except Exception as e:
+        # Fallback JSON slice
+        s, e2 = raw.find("{"), raw.rfind("}")
+        if s != -1 and e2 != -1 and e2 > s:
+            return json.loads(raw[s:e2+1])
+        raise RuntimeError(f"Manager returned non-JSON: {raw}") from e
+
+# ---------- Orchestration ----------
+
+def run_pipeline(file: str,
+                 oai_model: str = "gpt-4o-mini",
+                 gem_model: str = "gpt-4.1-nano",
+                 ds_model: str = "gpt-4.1-mini") -> Dict[str, Any]:
+    """
+    - Reads text (string or path) with load_input_text
+    - Calls three agents
+    - Saves their drafts as agent1.md / agent2.md / agent3.md
+    - Calls manager and returns its JSON (scores + winner)
+    """
+    source_text = load_input_text(file)
+
+    print("Generating case studies with three agents...")
+    print("Generating Agent 1 Output...")
+    a1 = call_openai_case_study(source_text, model=oai_model)
+    print("Generating Agent 2 Output...")
+    a2 = call_gemini_case_study(source_text, model=gem_model)
+    print("Generating Agent 3 Output...")
+    a3 = call_deepseek_case_study(source_text, model=ds_model)
+
+    pathlib.Path("agent1.md").write_text(a1, encoding="utf-8")
+    pathlib.Path("agent2.md").write_text(a2, encoding="utf-8")
+    pathlib.Path("agent3.md").write_text(a3, encoding="utf-8")
+    print("Saved agent outputs to agent1.md, agent2.md, agent3.md")
+
+    print("Manager evaluating...")
+    result = call_openai_manager(a1, a2, a3)
+    return result