How can I remove markdown formatting from SQL queries generated by SQLDatabaseSequentialChain in Langchain?

Below is one approach you can take to work around the issue:

1. Customize the Prompt or Postprocess Output
   The SQL query is being returned with markdown code blocks (e.g. “sql …”). This isn’t valid SQL for your database engine. To fix this, you can either modify the prompt given to the language model so that it outputs plain SQL without markdown formatting, or you can add a postprocessing step that strips out the markdown delimiters before executing the query.

2. Example: Using a Postprocessing Function
   You can define a helper function that removes the markdown symbols from the generated SQL query. For example:

   PYTHONCopied
   import re
   
   def remove_markdown(sql_query: str) -> str:
       # This pattern matches markdown code fences with an optional language identifier.
       cleaned = re.sub(r"```(?:\w+)?\s*(.*?)\s*```", r"\1", sql_query, flags=re.DOTALL)
       return cleaned.strip()

   Then, when you get your response from the chain, pass it through your function before sending it to SQLAlchemy:

   PYTHONCopied
   # Generate the response from the SequentialChain
   response = chain.invoke({"query": "How many user are there?"})
   
   # Remove markdown formatting from the SQL query
   cleaned_sql = remove_markdown(response)
   print("Clean SQL:", cleaned_sql)
   
   # Now you can execute cleaned_sql without the markdown formatting.

3. Alternatively, Customize the LLM Prompt
   If you prefer to have the LLM output plain SQL directly, update the prompt template used for the chain. You can instruct it explicitly (e.g., “output only the SQL query in plain text without any markdown or code block formatting”). When you customize the prompt passed into the chain’s factory function (like in create_sql_query_chain), your instructions might look like this:

   PYTHONCopied
   from langchain.prompts.prompt import PromptTemplate
   
   custom_template = """
   Given an input question, create a syntactically correct SQL query to run.
   Only output the SQL query in plain text (do not wrap it in markdown or code fences).
   
   Question: {input}
   SQL Query:"""
   prompt = PromptTemplate(
       input_variables=["input"], 
       template=custom_template
   )
   
   chain = SQLDatabaseSequentialChain.from_llm(
       llm=llm,
       database=db,
       verbose=True,
       prompt=prompt  # Use your custom prompt
   )
   
   response = chain.invoke({"query": "How many user are there?"})

   With this prompt, the model is less likely to output markdown markers along with your SQL query.

Both approaches solve the underlying issue: SQLAlchemy doesn’t understand markdown formatting. You can choose (or combine) the method that best suits your workflow.