Teradata MCP Server

# Chat Completion Tools Configuration # Provides tools for calling OpenAI-compatible LLM inference servers from Teradata chat_ai_mapreduce: type: prompt description: > Multi-step workflow to answer a high-level question using Teradata SQL and the chat_aggregatedCompleteChat tool. The agent first builds a Teradata query, then runs aggregated chat completion, and finally synthesizes a global answer. prompt: | You are going to answer the high-level question: "{question}". You must answer this question in three steps: Step 1 – Build the SQL query - Create a Teradata SQL query that selects the texts relevant to this high-level question. - The query must return a single column, renamed to "txt". - If it is reasonable for this question, filter the rows to keep only texts that are relevant to the question. - If such filtering is not clearly possible or meaningful, skip the filtering and explain why in your reasoning (but not in the SQL). - Remember you are querying a Teradata database, so use valid Teradata SQL syntax. - Unless the question explicitly states that there must be no sampling, add a `SAMPLE 1000` clause after any filtering to limit the number of rows. - Use the available tools to discover actual databases, tables, and columns before writing the final query. - Do not add a semicolon at the end of the SQL statement. - Do not use any other characters besides simple UTF-8 characters. - In your final output for this step, provide only the SQL query, nothing else. Step 2 – Run aggregated chat completion - Call the tool `chat_aggregatedCompleteChat`. - Pass the SQL from Step 1 as the `sql` parameter. - For the `system_message` parameter, construct a system prompt for the LLM that: - Focuses on a single text row at a time (do not ask about the whole dataset at once). - Guides the model to produce an answer that helps to answer the original high-level question, but from the perspective of just that one text. - Strongly enforces that the response for each text must be very short (no longer than 2–3 words). - Instructs the model to return an empty string if the text is not relevant to the high-level question. - May include examples of possible responses, but must not restrict the output to a fixed closed list. - Do not use any other characters besides simple UTF-8 characters. Step 3 – Synthesize the aggregated answer - Use the aggregated results from Step 2 (unique `response_txt` values and their counts) to produce a final, high-level answer to the original question. - When summarizing, remember that the responses in Step 2 were LLM-generated labels, so: - Different labels might refer to the same underlying reason or category (for example, "bad quality" vs "poor quality"). - Where appropriate, merge or interpret similar responses together. - Provide a clear, concise summary that explains the dominant patterns and insights that answer the high-level question. Follow these 3 steps exactly, and do not invent your own steps parameters: question: name: question description: "High-level question you want to answer using aggregated analysis over text data in Teradata. Explicitly state if sampling is allowed or not." required: true type_hint: str