Start wrapping chainladder to use as MCP tools

rag/server.py

@@ -8,7 +8,8 @@
 from langchain_ollama import OllamaEmbeddings
 from langchain.chat_models import init_chat_model
 from dotenv import load_dotenv
-from typing import List
+from typing import List, Dict, Any
+from triangle import triangle_build, dev_select, tail_constant, tail_curve, ibnr_estimate
 
 # Set up logging to stderr to avoid interfering with JSON-RPC over stdout
 logging.basicConfig(
@@ -68,6 +69,7 @@ def generate_answer(question: str, context_docs: List[Document]) -> str:
     response = llm.invoke(messages)
     return response.content
 
+
 @mcp.tool()
 def search_friedland_paper(prompt: str) -> str:
     """Search the Friedland actuarial paper for information"""
@@ -88,5 +90,125 @@ def search_both_papers(prompt: str) -> str:
     all_docs = friedland_docs + werner_modlin_docs
     return generate_answer(prompt, all_docs)
 
+@mcp.tool()
+def triangle_build_tool(input_data: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Build an actuarial triangle from input data using chainladder-python.
+
+    Required Input: { 
+        rows: [{origin, dev, value, segment?}], 
+        value_type: "cumulative"|"incremental", 
+        metric: "PaidLoss"|"ReportedLoss"|...,
+    }
+
+    Optional Input: {
+        exposure: [15000, 18000, 20000],  # Exposure per origin period
+        origin_col: "origin",             # Name of origin column in rows (default: 'origin')
+        dev_col: "dev",                   # Name of development column in rows (default: 'dev') 
+        value_col: "value"                # Name of value column in rows (default: 'value')
+    }
+
+    Output: { 
+        triangle_id, 
+        profile: {n_origin, n_dev, has_exposure}, 
+        warnings: [] 
+    }
+    """
+    return triangle_build(input_data)
+
+@mcp.tool()
+def dev_select_tool(input_data: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Perform development factor selection using chainladder.
+
+    Input: {
+        triangle_id: "uuid-string",
+        averaging: "volume"|"simple"|"median",
+        age_exclusions: [],
+        min_obs: 1,
+        tail_spec: null
+    }
+
+    Output: {
+        age_to_age: [],
+        LDF: [],
+        CDF: [],
+        tail_factor: 1.0,
+        diagnostics: {}
+    }
+    """
+    return dev_select(input_data)
+
+@mcp.tool()
+def tail_constant_tool(input_data: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Apply constant tail factor to triangle.
+
+    Input: {
+        triangle_id: "uuid-string",
+        tail_factor: 1.05
+    }
+
+    Output: {
+        tail_factor: 1.05,
+        ldf: [],
+        cdf: [],
+        diagnostics: {}
+    }
+    """
+    return tail_constant(input_data)
+
+@mcp.tool()
+def tail_curve_tool(input_data: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Apply curve-fitted tail to triangle.
+
+    Input: {
+        triangle_id: "uuid-string", 
+        extrap_periods: 100,
+        fit_period: null
+    }
+
+    Output: {
+        tail_factor: 1.03,
+        ldf: [],
+        cdf: [],
+        diagnostics: {}
+    }
+    """
+    return tail_curve(input_data)
+
+@mcp.tool()
+def ibnr_estimate_tool(input_data: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Calculate IBNR and Ultimate using various actuarial methods.
+
+    Input: {
+        triangle_id: "uuid-string",
+        method: "chainladder"|"bornhuetter_ferguson"|"benktander"|"expected_losses",
+        apriori: 0.75,  # Required for bornhuetter_ferguson, benktander, expected_losses
+        n_iters: 1,
+        trend: 0.0
+    }
+
+    Note: 
+    - chainladder method only requires triangle_id
+    - Other methods require exposure data in triangle_build
+
+    Output: {
+        ultimate: [],
+        ibnr: [],
+        latest_diagonal: [],
+        diagnostics: {}
+    }
+
+    TODO: Implement additional methods:
+    - case_outstanding: Requires incurred vs paid triangle structure
+    - cape_cod: Needs specific exposure data formatting  
+    - berquist_sherman: Requires additional parameter configuration
+    - frequency_severity: Custom implementation needed for the 3 variations(not in chainladder)
+    """
+    return ibnr_estimate(input_data)
+
 if __name__ == "__main__":
     mcp.run()