ChemiQuery – Natural Language to SQL

# User asks a question in plain English
question = "List all chemicals that are both carcinogenic and flammable"

# LangGraph state machine processes the request
state = {
    "question": question,
    "query": "",      # Will be filled by T5
    "result": "",     # Will be filled by SQL execution
    "answer": ""      # Final natural language response
}

# Execute the workflow
result = graph.invoke(state)
print(result["answer"])
# Output: "Found 12 chemicals matching your criteria:
#          1. Benzene (CAS: 71-43-2) - Location: Chemistry Building, Room 301
#          2. Formaldehyde (CAS: 50-00-0) - Location: Biology Lab, Room 105
#          ..."

Tables:
CREATE TABLE Chem_Chemical (
  ChemicalId INT PRIMARY KEY,
  Name NVARCHAR(255),
  CASNumber NVARCHAR(50),
  is_reviewed BIT,
  is_active BIT
);

CREATE TABLE Chem_Hazard (
  HazardId INT PRIMARY KEY,
  ChemicalId INT FOREIGN KEY,
  is_toxic BIT,
  is_carcinogen BIT,
  is_flammable BIT,
  is_corrosive BIT,
  NFPA_Health INT,
  NFPA_Flammability INT,
  -- 30+ additional hazard flags...
);

Question:
Find chemicals with NFPA health rating above 3 that haven't been reviewed

Answer:
SELECT c.Name, c.CASNumber, h.NFPA_Health
FROM Chem_Chemical c
JOIN Chem_Hazard h ON c.ChemicalId = h.ChemicalId
WHERE h.NFPA_Health > 3 AND c.is_reviewed = 0

from langgraph.graph import StateGraph, END
from typing import TypedDict

class QueryState(TypedDict):
    question: str
    query: str
    result: str
    answer: str

# Define the workflow graph
workflow = StateGraph(QueryState)

# Add nodes
workflow.add_node("run_finetuned_llm", generate_sql_with_t5)
workflow.add_node("execute_query", run_sql_on_database)
workflow.add_node("fall_back", retry_with_llama)
workflow.add_node("generate_answer", format_response)

# Define edges with conditional routing
workflow.set_entry_point("run_finetuned_llm")
workflow.add_edge("run_finetuned_llm", "execute_query")
workflow.add_conditional_edges(
    "execute_query",
    lambda x: "fall_back" if "error" in x["result"] else "generate_answer"
)
workflow.add_edge("fall_back", "execute_query")
workflow.add_edge("generate_answer", END)

graph = workflow.compile()

from transformers import TrainingArguments, Trainer

training_args = TrainingArguments(
    output_dir="./finetuned_model_2_epoch",
    num_train_epochs=2,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    learning_rate=5e-3,
    weight_decay=0.01,
    evaluation_strategy="steps",
    eval_steps=500,
    logging_steps=50,
    save_strategy="epoch",
    load_best_model_at_end=True,
    bf16=True,  # bfloat16 for memory efficiency
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_train,
    eval_dataset=tokenized_val,
    compute_metrics=compute_rouge_metrics,
)

# Train and evaluate
trainer.train()
# Results: ROUGE-1: 0.996, ROUGE-L: 0.996