aminghadersohi commented on code in PR #37200:
URL: https://github.com/apache/superset/pull/37200#discussion_r2775637245
##########
superset/mcp_service/mcp_config.py:
##########
@@ -155,6 +155,50 @@
],
}
+# =============================================================================
+# MCP Response Size Guard Configuration
+# =============================================================================
+#
+# Overview:
+# ---------
+# The Response Size Guard prevents oversized responses from overwhelming LLM
+# clients (e.g., Claude Desktop). When a tool response exceeds the token limit,
+# it returns a helpful error with suggestions for reducing the response size.
+#
+# How it works:
+# -------------
+# 1. After a tool executes, the middleware estimates the response's token count
+# 2. If the response exceeds the configured limit, it blocks the response
+# 3. Instead, it returns an error message with smart suggestions:
+# - Reduce page_size/limit
+# - Use select_columns to exclude large fields
+# - Add filters to narrow results
+# - Tool-specific recommendations
+#
+# Configuration:
+# --------------
+# - enabled: Toggle the guard on/off (default: True)
+# - token_limit: Maximum estimated tokens per response (default: 25,000)
+# - excluded_tools: Tools to skip checking (e.g., streaming tools)
+# - warn_threshold_pct: Log warnings above this % of limit (default: 80%)
+#
+# Token Estimation:
+# -----------------
+# Uses character-based heuristic (~3.5 chars per token for JSON).
+# This is intentionally conservative to avoid underestimating.
+# =============================================================================
+MCP_RESPONSE_SIZE_CONFIG: Dict[str, Any] = {
+ "enabled": True, # Enabled by default to protect LLM clients
+ "token_limit": 25000, # ~25k tokens prevents overwhelming LLM context
windows
Review Comment:
Done. Changed to `25_000` and moved to `constants.py` with
`DEFAULT_TOKEN_LIMIT`.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
