feat: improve tabular SK large result handling

application/single_app/config.py

@@ -95,7 +95,7 @@
 EXECUTOR_TYPE = 'thread'
 EXECUTOR_MAX_WORKERS = 30
 SESSION_TYPE = 'filesystem'
-VERSION = "0.242.066"
+VERSION = "0.242.068"
 
 SESSION_COOKIE_SAMESITE = os.getenv('SESSION_COOKIE_SAMESITE', 'Lax')
 SESSION_COOKIE_HTTPONLY = os.getenv('SESSION_COOKIE_HTTPONLY', 'true').lower() != 'false'

application/single_app/route_backend_chats.py

@@ -41,6 +41,7 @@
 import queue
 import re
 import requests
+import time
 import traceback
 from urllib.parse import urlparse
 import threading
@@ -1098,6 +1099,8 @@ def _strip_agent_citation_artifact_refs(agent_citations):
 TABULAR_RELATED_DOCUMENT_MAX_MATCHES_PER_ROW = 3
 TABULAR_RELATED_DOCUMENT_MAX_SUMMARY_ROWS = 8
 TABULAR_RELATED_DOCUMENT_MAX_EXCERPT_CHARS = 500
+TABULAR_SK_ANALYSIS_MAX_CHARS = 100000
+TABULAR_COMPUTED_RESULTS_HANDOFF_MAX_CHARS = 100000
 TABULAR_GENERATED_OUTPUT_INTERNAL_ROW_FIELDS = {
  '_matched_columns',
  '_matched_values',
@@ -3661,8 +3664,13 @@ def build_tabular_related_document_evidence_summary(invocations):
 def build_tabular_computed_results_system_message(source_label, tabular_analysis, related_document_evidence_summary=''):
  """Build the outer-model handoff message for successful tabular analysis."""
  rendered_analysis = str(tabular_analysis or '').strip()
- max_handoff_chars = 24000
+ max_handoff_chars = TABULAR_COMPUTED_RESULTS_HANDOFF_MAX_CHARS
  if len(rendered_analysis) > max_handoff_chars:
+ original_length = len(rendered_analysis)
+ log_event(
+ f"[Tabular SK Analysis] Computed results handoff truncated from {original_length} to {max_handoff_chars} chars",
+ level=logging.WARNING,
+ )
  rendered_analysis = (
  rendered_analysis[:max_handoff_chars]
  + "\n[Computed results handoff truncated for prompt budget.]"
@@ -9376,6 +9384,7 @@ def build_system_prompt(force_tool_use=False, tool_error_messages=None,
  return None
 
  for attempt_number in range(1, 4):
+ attempt_started_at = time.monotonic()
  force_tool_use = attempt_number > 1 or (attempt_number == 1 and analysis_requires_immediate_tool_choice)
  if callable(thought_callback) and attempt_number > 1:
  await emit_tabular_analysis_lifecycle_thought(
@@ -9494,13 +9503,24 @@ def build_system_prompt(force_tool_use=False, tool_error_messages=None,
 
  if result and result[0].content:
  analysis = result[0].content.strip()
- if len(analysis) > 20000:
- analysis = analysis[:20000] + "\n[Analysis truncated]"
+ if len(analysis) > TABULAR_SK_ANALYSIS_MAX_CHARS:
+ original_analysis_length = len(analysis)
+ log_event(
+ f"[Tabular SK Analysis] Attempt {attempt_number} analysis text truncated from {original_analysis_length} to {TABULAR_SK_ANALYSIS_MAX_CHARS} chars",
+ level=logging.WARNING,
+ )
+ analysis = analysis[:TABULAR_SK_ANALYSIS_MAX_CHARS] + "\n[Analysis truncated]"
+ attempt_elapsed_ms = int((time.monotonic() - attempt_started_at) * 1000)
 
  if schema_summary_mode:
  if successful_schema_summary_invocations:
  log_event(
  f"[Tabular SK Analysis] Schema summary complete via {len(successful_schema_summary_invocations)} workbook tool call(s) on attempt {attempt_number}",
+ extra={
+ 'attempt_number': attempt_number,
+ 'elapsed_ms': attempt_elapsed_ms,
+ 'analysis_length': len(analysis),
+ },
  level=logging.INFO,
  )
  return analysis
@@ -9601,6 +9621,11 @@ def build_system_prompt(force_tool_use=False, tool_error_messages=None,
  previous_execution_gap_messages = []
  log_event(
  f"[Tabular SK Analysis] Analysis complete via {len(successful_analytical_invocations)} analytical tool call(s) on attempt {attempt_number}",
+ extra={
+ 'attempt_number': attempt_number,
+ 'elapsed_ms': attempt_elapsed_ms,
+ 'analysis_length': len(analysis),
+ },
  level=logging.INFO
  )
  return analysis