LLM Refinement Layer

After executing the SQL query and retrieving the raw results, Billx-Agent can optionally send the data back to the LLM for natural language summarization. This helps convert technical output into user-friendly answers.

---

🎯 What It Does

• Takes the raw SQL output (rows + columns)

• Rephrases it into a plain-English summary

• Optionally applies context, units, or comparisons

---

💡 Why It’s Useful

Raw SQL results like this:

[{ "sum": 24320 }]

…can be refined to:

"Total sales for Q1 were $24,320"

This is especially helpful when:

• Embedding results in dashboards or reports

• Reading the answer aloud in /audio-chat

• Serving non-technical users

---

🔁 When Is It Triggered?

• Automatically enabled on /chat and /audio-chat

• Powered by the same Gemini LLM engine used for SQL generation

• Uses the original prompt + result to generate a final answer

---

📋 Example Refinement Call

Input to LLM:

{
  "prompt": "How many orders were placed last month?",
  "result": [{ "count": 218 }]
}

LLM Output:

"There were 218 orders placed in the last month."

---

🧠 Prompt + Result Fusion

The LLM receives:

• The original prompt

• The SQL query

• The raw result

• Optionally: the tool name and database context

This results in responses that are accurate, conversational, and safe.

---

🧑‍🏫 You can disable refinement if you want pure SQL results only (e.g. analytics dashboards).