docoreai 2.3.7__tar.gz → 2.3.9__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {docoreai-2.3.7 → docoreai-2.3.9}/PKG-INFO +9 -4
- {docoreai-2.3.7 → docoreai-2.3.9}/README.md +5 -2
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/auto_patch.py +276 -16
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/cli.py +33 -1
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/config/db.py +112 -6
- docoreai-2.3.9/docore_ai/instrumentation/adapters/autogen.py +224 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/adapters/langchain.py +2 -2
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/adapters/openai_agents.py +8 -2
- docoreai-2.3.9/docore_ai/instrumentation/agentic_cost/__init__.py +2 -0
- docoreai-2.3.9/docore_ai/instrumentation/agentic_cost/embedder.py +69 -0
- docoreai-2.3.9/docore_ai/instrumentation/agentic_cost/prompt_compression.py +193 -0
- docoreai-2.3.9/docore_ai/instrumentation/agentic_cost/tool_pruning_cache.py +818 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/framework_registry.py +1 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/resolver.py +10 -0
- docoreai-2.3.9/docore_ai/reasoning_capture.py +123 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/telemetry_local.py +88 -5
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/telemetry_wp.py +30 -1
- docoreai-2.3.9/docore_ai/telemetry/token_utils.py +92 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/agentic_predictor.py +12 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/phase2_config_tab.py +6 -2
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/model_patch_targets_sync.py +13 -10
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/model_registry_sync.py +123 -49
- {docoreai-2.3.7 → docoreai-2.3.9}/docoreai.egg-info/SOURCES.txt +6 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/setup.py +5 -2
- docoreai-2.3.7/docore_ai/telemetry/token_utils.py +0 -46
- {docoreai-2.3.7 → docoreai-2.3.9}/CONTRIBUTING-TESTERS.md +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/DEPLOYMENT.md +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/LICENSE.md +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/MANIFEST.in +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/api/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/api/main.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/config/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/config/init_db_async.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/fusion_log_server.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/init_engine.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/adapters/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/adapters/base.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/adapters/crewai.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/instrumentation/context_manager.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/model.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/patch_manager.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/patcher.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/prompt_updates/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/prompt_updates/check_prompt_update.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/prompt_updates/prompt_loader.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/providers/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/providers/bedrock_adapter.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/providers/ollama.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/proxy.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/pth_installer.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/pth_installer_new.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/runtime/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/runtime/settings.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/comparator/base.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/comparator/llama_cpp_comparator.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/core/alert_emitter.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/core/confidence_engine.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/core/derived_gate.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/core/governance_dispatcher.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/core/policy.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/detectors/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/detectors/content_policy.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/detectors/pii_detector.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/export/exporter_map.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/export/sdk_hooks.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/lineage/fingerprint.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/lineage/lifecycle.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/lineage/tracker.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/semantic/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/semantic/matcher.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/semantic/sketch.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/semantic/thresholds.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/state/cache.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/state/context.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/state/ttl.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/structure/detector.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/structure/heuristics.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/security/types.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/config.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/dispatcher.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/safe_sender.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/safe_sender1.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/token_auth.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/tracker.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/telemetry/usage_limits.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/test_server.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/ab_test_controller.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/auto_retrain_manager.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/auto_trainer.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/budget_manager.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/budget_recommendation.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/cli_integration.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/config.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/daily_pacing_engine.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/data_extraction.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/drift_detector.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/drift_monitor_service.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/evaluation.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/feature_engineering.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/hourly_patterns.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/local_budget_checker.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/monitor.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/performance_tracker.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/phase2_config.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/predict.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/runtime_predictor.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/session_tracker.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/soft_limit_injector.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/spend_forecaster.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/tests/test_fetch_budget_api.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/tests/test_phase1.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/tests/test_phase2.1_complete.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/tests/test_phase2.2_live.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/tests/test_phase2_mock.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/tests/test_quick.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/token_governor.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/tokenization_learner.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/train/train_model.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/charts/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/charts/cost_chart.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/charts/daily_spend_chart.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/charts/usage_requests_chart.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/deployment_map_tab.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/governance_event_log.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/live_tab.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/money_burn_tab.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/settings_header.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/text_input.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/toggle.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/components/usage_volume_tab.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/launcher.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/ui/main_window.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/EXTINCT-model_registry.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/__init__.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/budget_sync.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/csv_logger.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/helpers.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/logger.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/metric_aggregator.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/run_streamlit_with_shutdown.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/docore_ai/utils/show_app.py +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/requirements.txt +0 -0
- {docoreai-2.3.7 → docoreai-2.3.9}/setup.cfg +0 -0
|
@@ -1,9 +1,9 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: docoreai
|
|
3
|
-
Version: 2.3.
|
|
3
|
+
Version: 2.3.9
|
|
4
4
|
Summary: Self-tuning AI Cost Governance, Built Privacy-First.
|
|
5
5
|
Home-page: https://docoreai.com
|
|
6
|
-
Author: Saji John
|
|
6
|
+
Author: Saji John
|
|
7
7
|
Author-email: saji.john@docoreai.com
|
|
8
8
|
License: CC BY-NC-ND 4.0
|
|
9
9
|
Project-URL: Homepage, https://docoreai.com
|
|
@@ -53,6 +53,8 @@ Requires-Dist: charset-normalizer>=3.0.0
|
|
|
53
53
|
Requires-Dist: google-genai
|
|
54
54
|
Requires-Dist: prophet>=1.1
|
|
55
55
|
Requires-Dist: cloud-detect>=0.0.15
|
|
56
|
+
Requires-Dist: sentence-transformers
|
|
57
|
+
Requires-Dist: llmlingua
|
|
56
58
|
Dynamic: author
|
|
57
59
|
Dynamic: author-email
|
|
58
60
|
Dynamic: classifier
|
|
@@ -105,7 +107,7 @@ LLM costs do not self-regulate. Without active budget control, a traffic surge a
|
|
|
105
107
|
|
|
106
108
|
DoCoreAI sits between your application and any LLM provider — OpenAI, Anthropic, Google, Groq, AWS Bedrock — and acts as an autonomous cost and governance layer. It instruments every LLM call your application makes: tracking cost, tokens, latency, and governance decisions — without ever capturing prompt content, response content, or user data.
|
|
107
109
|
|
|
108
|
-
It learns your actual usage patterns. It predicts what each request genuinely needs. It replaces the wasteful default token ceiling with a precise ML prediction. It paces your budget intelligently across the full day and across the full week, accounting for day-of-week traffic patterns. When its own accuracy degrades, it retrains automatically. Your prompts never leave your network.
|
|
110
|
+
It learns your actual usage patterns. It predicts what each request genuinely needs. It replaces the wasteful default token ceiling with a precise ML prediction. It paces your budget intelligently across the full day and across the full week, accounting for day-of-week traffic patterns. It trims redundant context before requests leave your environment — shrinking token counts without changing what the model sees. When its own accuracy degrades, it retrains automatically. Your prompts never leave your network.
|
|
109
111
|
|
|
110
112
|
**This is not observability.** Observability shows you what happened after the bill arrives. DoCoreAI predicts, paces, and controls spend before overruns occur — autonomously, on every request.
|
|
111
113
|
|
|
@@ -189,6 +191,8 @@ The default `max_tokens` on most LLM calls is set to 2,000 or higher — a safe
|
|
|
189
191
|
|
|
190
192
|
At 100,000 daily calls, that single optimisation saves approximately **$174/day — $63,510/year** per deployment. Before pacing, soft limits, or session governance are applied.
|
|
191
193
|
|
|
194
|
+
A second lever reduces input token costs. On agentic turns, only the tool definitions the request actually needs are sent — large tool schemas passed on every call are a common source of silent prompt bloat. On multi-step agent sessions, long tool-call results are compressed before re-entering the context window. Both mechanisms are semantic — meaning is preserved, content is not truncated. Combined with output prediction and pacing, the 40–70% range reflects deployments where all three levers are active.
|
|
195
|
+
|
|
192
196
|
---
|
|
193
197
|
|
|
194
198
|
## Quick Start
|
|
@@ -312,7 +316,7 @@ Works out of the box with any combination. No provider-specific configuration re
|
|
|
312
316
|
- **Groq** — Llama, Mixtral, and all Groq-hosted models
|
|
313
317
|
- **Ollama** — Local model deployments
|
|
314
318
|
|
|
315
|
-
**Multi-LLM orchestration:** LangChain, CrewAI (v1.x — LiteLLM-backed), OpenAI Agents SDK, and custom orchestration frameworks are all supported transparently. DoCoreAI patches at the SDK level — whatever framework sits above it, the governance fires on every call.
|
|
319
|
+
**Multi-LLM orchestration:** LangChain, CrewAI (v1.x — LiteLLM-backed), OpenAI Agents SDK, and custom orchestration frameworks are all supported transparently. DoCoreAI patches at the SDK level — whatever framework sits above it, the governance fires on every call.
|
|
316
320
|
|
|
317
321
|
---
|
|
318
322
|
|
|
@@ -337,6 +341,7 @@ Works out of the box with any combination. No provider-specific configuration re
|
|
|
337
341
|
| Executive reporting suite | ✗ | ✓ 7 reports |
|
|
338
342
|
| Compliance-ready architecture | ✗ | ✓ |
|
|
339
343
|
| Self-improving prediction model | ✗ | ✓ Auto-retrain + A/B test |
|
|
344
|
+
| Context optimization (tool schema + result compression) | ✗ | ✓ |
|
|
340
345
|
| Closed source / IP protected | ✗ | ✓ from v2.0 |
|
|
341
346
|
| SOC2 | Varies | In progress |
|
|
342
347
|
|
|
@@ -36,7 +36,7 @@ LLM costs do not self-regulate. Without active budget control, a traffic surge a
|
|
|
36
36
|
|
|
37
37
|
DoCoreAI sits between your application and any LLM provider — OpenAI, Anthropic, Google, Groq, AWS Bedrock — and acts as an autonomous cost and governance layer. It instruments every LLM call your application makes: tracking cost, tokens, latency, and governance decisions — without ever capturing prompt content, response content, or user data.
|
|
38
38
|
|
|
39
|
-
It learns your actual usage patterns. It predicts what each request genuinely needs. It replaces the wasteful default token ceiling with a precise ML prediction. It paces your budget intelligently across the full day and across the full week, accounting for day-of-week traffic patterns. When its own accuracy degrades, it retrains automatically. Your prompts never leave your network.
|
|
39
|
+
It learns your actual usage patterns. It predicts what each request genuinely needs. It replaces the wasteful default token ceiling with a precise ML prediction. It paces your budget intelligently across the full day and across the full week, accounting for day-of-week traffic patterns. It trims redundant context before requests leave your environment — shrinking token counts without changing what the model sees. When its own accuracy degrades, it retrains automatically. Your prompts never leave your network.
|
|
40
40
|
|
|
41
41
|
**This is not observability.** Observability shows you what happened after the bill arrives. DoCoreAI predicts, paces, and controls spend before overruns occur — autonomously, on every request.
|
|
42
42
|
|
|
@@ -120,6 +120,8 @@ The default `max_tokens` on most LLM calls is set to 2,000 or higher — a safe
|
|
|
120
120
|
|
|
121
121
|
At 100,000 daily calls, that single optimisation saves approximately **$174/day — $63,510/year** per deployment. Before pacing, soft limits, or session governance are applied.
|
|
122
122
|
|
|
123
|
+
A second lever reduces input token costs. On agentic turns, only the tool definitions the request actually needs are sent — large tool schemas passed on every call are a common source of silent prompt bloat. On multi-step agent sessions, long tool-call results are compressed before re-entering the context window. Both mechanisms are semantic — meaning is preserved, content is not truncated. Combined with output prediction and pacing, the 40–70% range reflects deployments where all three levers are active.
|
|
124
|
+
|
|
123
125
|
---
|
|
124
126
|
|
|
125
127
|
## Quick Start
|
|
@@ -243,7 +245,7 @@ Works out of the box with any combination. No provider-specific configuration re
|
|
|
243
245
|
- **Groq** — Llama, Mixtral, and all Groq-hosted models
|
|
244
246
|
- **Ollama** — Local model deployments
|
|
245
247
|
|
|
246
|
-
**Multi-LLM orchestration:** LangChain, CrewAI (v1.x — LiteLLM-backed), OpenAI Agents SDK, and custom orchestration frameworks are all supported transparently. DoCoreAI patches at the SDK level — whatever framework sits above it, the governance fires on every call.
|
|
248
|
+
**Multi-LLM orchestration:** LangChain, CrewAI (v1.x — LiteLLM-backed), OpenAI Agents SDK, and custom orchestration frameworks are all supported transparently. DoCoreAI patches at the SDK level — whatever framework sits above it, the governance fires on every call.
|
|
247
249
|
|
|
248
250
|
---
|
|
249
251
|
|
|
@@ -268,6 +270,7 @@ Works out of the box with any combination. No provider-specific configuration re
|
|
|
268
270
|
| Executive reporting suite | ✗ | ✓ 7 reports |
|
|
269
271
|
| Compliance-ready architecture | ✗ | ✓ |
|
|
270
272
|
| Self-improving prediction model | ✗ | ✓ Auto-retrain + A/B test |
|
|
273
|
+
| Context optimization (tool schema + result compression) | ✗ | ✓ |
|
|
271
274
|
| Closed source / IP protected | ✗ | ✓ from v2.0 |
|
|
272
275
|
| SOC2 | Varies | In progress |
|
|
273
276
|
|
|
@@ -208,6 +208,13 @@ def inject_system_prompt(provider: str, kwargs: dict, model_name: str) -> None:
|
|
|
208
208
|
# if not active_prompt:
|
|
209
209
|
# return
|
|
210
210
|
|
|
211
|
+
# Skip soft limit injection when tools are present — word-count guidance
|
|
212
|
+
# is meaningless for tool-calling turns (mirrors max_tokens skip logic)
|
|
213
|
+
if kwargs.get('tools'):
|
|
214
|
+
dprint("⏭️ Soft limit injection skipped — tools present in request")
|
|
215
|
+
kwargs['_docoreai_soft_limit_tokens_applied'] = 0
|
|
216
|
+
return
|
|
217
|
+
|
|
211
218
|
# Get soft limit text — generated dynamically from governance ceiling
|
|
212
219
|
soft_limit_tokens = (
|
|
213
220
|
kwargs.get('_docoreai_injected_max_tokens') or
|
|
@@ -1086,7 +1093,59 @@ def prepare_and_log_telemetry(args, kwargs, response, provider, in_learning_wind
|
|
|
1086
1093
|
dprint(f"⚠️ Response metrics logging failed: {response_log_error}")
|
|
1087
1094
|
|
|
1088
1095
|
|
|
1096
|
+
# Responses API path: system prompt is in 'instructions' kwarg, not messages.
|
|
1097
|
+
# _docoreai_system_prompt_tokens will be 0 in that case — fall back to instructions.
|
|
1098
|
+
_sys_tokens = kwargs.get('_docoreai_system_prompt_tokens', 0)
|
|
1099
|
+
if not _sys_tokens:
|
|
1100
|
+
_instr = kwargs.get('instructions', '') or ''
|
|
1101
|
+
if _instr:
|
|
1102
|
+
_sys_tokens = len(_instr) // 4
|
|
1103
|
+
dprint(f"📊 system_prompt_tokens from 'instructions' kwarg: {_sys_tokens}")
|
|
1089
1104
|
|
|
1105
|
+
# ── Input Savings Calculation ──────────────────────────────────────────
|
|
1106
|
+
# Two separate optimizations reduce LLM INPUT tokens. Each is measured
|
|
1107
|
+
# from values stashed in kwargs during the pre-call phase.
|
|
1108
|
+
#
|
|
1109
|
+
# 1. TOOL PRUNING (tool_pruning_cache.py → check_cache())
|
|
1110
|
+
# Removes entire tool DEFINITIONS from the `tools` parameter.
|
|
1111
|
+
# Values: _docoreai_tools_prompt_length (chars, original definitions)
|
|
1112
|
+
# _docoreai_tools_prompt_length_after_prune (chars, after pruning)
|
|
1113
|
+
# _docoreai_tools_prompt_tokens (token count, original)
|
|
1114
|
+
# Method: byte-ratio on the definition JSON — same content type, accurate.
|
|
1115
|
+
#
|
|
1116
|
+
# 2. PROMPT COMPRESSION (prompt_compression.py → compress_agentic_tool_results())
|
|
1117
|
+
# Reduces role='tool' RESULT MESSAGES in the `messages` parameter.
|
|
1118
|
+
# Values: _docoreai_tool_content_original (chars, original tool results)
|
|
1119
|
+
# _docoreai_tool_content_after_comp (chars, after LLMLingua-2)
|
|
1120
|
+
# Method: fixed 4 chars/token — tool results are plain text, not JSON.
|
|
1121
|
+
# DO NOT use tools_prompt_tokens ratio here: that ratio is for
|
|
1122
|
+
# structured definition JSON (~5–7 chars/tok), not plain text.
|
|
1123
|
+
# ──────────────────────────────────────────────────────────────────────
|
|
1124
|
+
_tel_tpt = kwargs.get('_docoreai_tools_prompt_tokens', 0) or 0
|
|
1125
|
+
_tel_tpl = kwargs.get('_docoreai_tools_prompt_length', 0) or 0
|
|
1126
|
+
_tel_pruned = kwargs.get('_docoreai_tools_prompt_length_after_prune', 0) or 0
|
|
1127
|
+
_tel_tco = kwargs.get('_docoreai_tool_content_original', 0) or 0
|
|
1128
|
+
_tel_tca = kwargs.get('_docoreai_tool_content_after_comp', 0) or 0
|
|
1129
|
+
|
|
1130
|
+
# 1. TOOL PRUNING savings — byte-ratio on tool definition JSON
|
|
1131
|
+
_tok_prune = 0
|
|
1132
|
+
if _tel_tpl > 0 and _tel_pruned < _tel_tpl and _tel_tpt > 0:
|
|
1133
|
+
_tok_prune = max(0, round(_tel_tpt * (1.0 - _tel_pruned / _tel_tpl)))
|
|
1134
|
+
|
|
1135
|
+
# 2. PROMPT COMPRESSION savings — 4 chars/token for plain tool result text
|
|
1136
|
+
_tok_comp = 0
|
|
1137
|
+
if _tel_tco > 0 and _tel_tca < _tel_tco:
|
|
1138
|
+
_tok_comp = max(0, round((_tel_tco - _tel_tca) / 4.0))
|
|
1139
|
+
|
|
1140
|
+
# Combined input tokens saved from both optimizations this call
|
|
1141
|
+
_tokens_saved_input = _tok_prune + _tok_comp
|
|
1142
|
+
|
|
1143
|
+
# Cost saved using price_per_1k_prompt from model_registry at call time.
|
|
1144
|
+
# Stored so historical rows remain accurate if pricing changes later.
|
|
1145
|
+
dprint(f"💰 Input savings — prune: {_tok_prune} tok (tool defs) | "
|
|
1146
|
+
f"comp: {_tok_comp} tok (tool results) | "
|
|
1147
|
+
f"total: {_tokens_saved_input} tok")
|
|
1148
|
+
# ──────────────────────────────────────────────────────────────────────
|
|
1090
1149
|
payload = prepare_telemetry_payload(
|
|
1091
1150
|
user_id=state.get("user_id"),
|
|
1092
1151
|
user_content=user_content,
|
|
@@ -1104,13 +1163,18 @@ def prepare_and_log_telemetry(args, kwargs, response, provider, in_learning_wind
|
|
|
1104
1163
|
mode='T' if in_learning_window else 'L',
|
|
1105
1164
|
session_id = kwargs.get('_docoreai_session_id', None),
|
|
1106
1165
|
session_type = kwargs.get('_docoreai_session_type'),
|
|
1107
|
-
system_prompt_tokens = kwargs.get('_docoreai_system_prompt_tokens', 0),
|
|
1166
|
+
system_prompt_tokens = _sys_tokens, #kwargs.get('_docoreai_system_prompt_tokens', 0),
|
|
1108
1167
|
tools_prompt_tokens = kwargs.get('_docoreai_tools_prompt_tokens', 0),
|
|
1109
1168
|
tools_prompt_length = kwargs.get('_docoreai_tools_prompt_length', 0),
|
|
1110
1169
|
temperature = kwargs.get('temperature', -1),
|
|
1111
1170
|
client_prompt_id = kwargs.get('_docoreai_client_prompt_id'),
|
|
1112
1171
|
soft_limit_tokens_applied=kwargs.get('_docoreai_soft_limit_tokens_applied',0),
|
|
1113
|
-
soft_limit_target_pct_reduced=float(get_phase2_config('soft_limit', 'target_reduction_pct', 0.0))
|
|
1172
|
+
soft_limit_target_pct_reduced=float(get_phase2_config('soft_limit', 'target_reduction_pct', 0.0)),
|
|
1173
|
+
tool_content_original = kwargs.get('_docoreai_tool_content_original', 0),
|
|
1174
|
+
tool_content_after_comp = kwargs.get('_docoreai_tool_content_after_comp', 0),
|
|
1175
|
+
tools_prompt_length_after_prune= kwargs.get('_docoreai_tools_prompt_length_after_prune', 0),
|
|
1176
|
+
tokens_saved_input = _tokens_saved_input, # total input tokens avoided
|
|
1177
|
+
|
|
1114
1178
|
)
|
|
1115
1179
|
|
|
1116
1180
|
# Attach agentic classification (set by make_llm_wrapper via kwargs stash).
|
|
@@ -1122,6 +1186,8 @@ def prepare_and_log_telemetry(args, kwargs, response, provider, in_learning_wind
|
|
|
1122
1186
|
payload["call_type"] = _call_type
|
|
1123
1187
|
payload["framework_name"] = kwargs.get("_docoreai_framework_name") or None
|
|
1124
1188
|
payload["agent_name"] = kwargs.get("_docoreai_agent_name") or None
|
|
1189
|
+
payload["reasoning_effort"] = kwargs.get("_docoreai_reasoning_effort") or None
|
|
1190
|
+
payload["reasoning_tokens"] = kwargs.get("_docoreai_reasoning_tokens") or None
|
|
1125
1191
|
|
|
1126
1192
|
|
|
1127
1193
|
log_telemetry_to_wordpress(success=True, telemetry_data=payload)
|
|
@@ -1157,10 +1223,31 @@ def _get_max_tokens_param(model_name: str) -> str | None:
|
|
|
1157
1223
|
pass
|
|
1158
1224
|
return None # not in registry — caller uses 'max_tokens' fallback
|
|
1159
1225
|
# ------------------------------------------------------------
|
|
1226
|
+
# REASONING MODEL GATE
|
|
1227
|
+
# ------------------------------------------------------------
|
|
1228
|
+
from functools import lru_cache
|
|
1229
|
+
|
|
1230
|
+
@lru_cache(maxsize=128)
|
|
1231
|
+
def _is_reasoning_capable(model_name: str) -> bool:
|
|
1232
|
+
"""Return True only for models with reasoning_capable=1 in model_registry."""
|
|
1233
|
+
try:
|
|
1234
|
+
from docore_ai.config.db import get_readonly_connection
|
|
1235
|
+
conn = get_readonly_connection()
|
|
1236
|
+
row = conn.execute(
|
|
1237
|
+
"SELECT reasoning_capable FROM wp_docoreai_model_registry "
|
|
1238
|
+
"WHERE model_name = ? AND status = 'active' LIMIT 1",
|
|
1239
|
+
(model_name,)
|
|
1240
|
+
).fetchone()
|
|
1241
|
+
conn.close()
|
|
1242
|
+
return bool(row and row[0])
|
|
1243
|
+
except Exception:
|
|
1244
|
+
return False
|
|
1245
|
+
|
|
1246
|
+
# ------------------------------------------------------------
|
|
1160
1247
|
# UNIVERSAL WRAPPER (provider bound at patch time)
|
|
1161
1248
|
# + GenAI Runtime Governance (Phase-1)
|
|
1162
1249
|
# ------------------------------------------------------------
|
|
1163
|
-
def make_llm_wrapper(provider, config_provider=None, input_field='messages', output_field='choices[0].message.content'):
|
|
1250
|
+
def make_llm_wrapper(provider, config_provider=None, input_field='messages', output_field='choices[0].message.content', reasoning_extractor=None):
|
|
1164
1251
|
def llm_wrapper(wrapped, instance, args, kwargs):
|
|
1165
1252
|
kwargs['_docoreai_input_field'] = input_field
|
|
1166
1253
|
kwargs['_docoreai_output_field'] = output_field
|
|
@@ -1490,7 +1577,7 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
1490
1577
|
prompt_tokens = len(full_prompt_text) // 4
|
|
1491
1578
|
prompt_length = len(full_prompt_text)
|
|
1492
1579
|
system_prompt_tokens = len(system_prompt_text) // 4
|
|
1493
|
-
|
|
1580
|
+
|
|
1494
1581
|
|
|
1495
1582
|
# Count token cost of tool schemas attached to this request
|
|
1496
1583
|
_tools_in_request = kwargs.get('tools', []) or []
|
|
@@ -1499,7 +1586,15 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
1499
1586
|
import json as _json_tools
|
|
1500
1587
|
_tools_json_str = _json_tools.dumps(_tools_in_request)
|
|
1501
1588
|
tools_prompt_length = len(_tools_json_str)
|
|
1502
|
-
|
|
1589
|
+
try:
|
|
1590
|
+
from docore_ai.telemetry.token_utils import _resolve_tokenizer
|
|
1591
|
+
_tok_mode, _tok_enc, _ = _resolve_tokenizer(current_model)
|
|
1592
|
+
if _tok_mode == 'tiktoken' and _tok_enc is not None:
|
|
1593
|
+
tools_prompt_tokens = len(_tok_enc.encode(_tools_json_str))
|
|
1594
|
+
else:
|
|
1595
|
+
tools_prompt_tokens = tools_prompt_length // 4
|
|
1596
|
+
except Exception:
|
|
1597
|
+
tools_prompt_tokens = tools_prompt_length // 4
|
|
1503
1598
|
except Exception:
|
|
1504
1599
|
tools_prompt_length = 0
|
|
1505
1600
|
tools_prompt_tokens = 0
|
|
@@ -2089,9 +2184,32 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
2089
2184
|
# (called before STEP 4 so _docoreai_injected_max_tokens not yet set)
|
|
2090
2185
|
if governance_max_tokens is not None:
|
|
2091
2186
|
kwargs['_docoreai_governance_max_tokens'] = governance_max_tokens
|
|
2092
|
-
#
|
|
2093
|
-
|
|
2187
|
+
# Skip soft limit injection when tools are present — word-count
|
|
2188
|
+
# guidance is meaningless for tool-calling turns.
|
|
2189
|
+
if not kwargs.get('tools'):
|
|
2190
|
+
inject_system_prompt(provider, kwargs, kwargs.get("model") or "unknown")
|
|
2191
|
+
else:
|
|
2192
|
+
dprint("⏭️ Soft limit injection skipped — tools present in request")
|
|
2193
|
+
kwargs['_docoreai_soft_limit_tokens_applied'] = 0
|
|
2094
2194
|
|
|
2195
|
+
# Recompute system_prompt_tokens after injection so the telemetry
|
|
2196
|
+
# captures the injected text that just landed in messages/instructions/system
|
|
2197
|
+
try:
|
|
2198
|
+
_input_field_for_sys = kwargs.get('_docoreai_input_field', 'messages')
|
|
2199
|
+
_msgs_after_inject = kwargs.get(_input_field_for_sys, []) or []
|
|
2200
|
+
_sys_text_after = " ".join([
|
|
2201
|
+
m.get('content', '') for m in _msgs_after_inject
|
|
2202
|
+
if isinstance(m, dict) and m.get('role') == 'system'
|
|
2203
|
+
])
|
|
2204
|
+
# Anthropic kwarg path
|
|
2205
|
+
_anth_sys = kwargs.get('system', '') or ''
|
|
2206
|
+
if isinstance(_anth_sys, list):
|
|
2207
|
+
_anth_sys = ' '.join(b.get('text', '') for b in _anth_sys if isinstance(b, dict))
|
|
2208
|
+
if _anth_sys:
|
|
2209
|
+
_sys_text_after = _anth_sys + ' ' + _sys_text_after
|
|
2210
|
+
kwargs['_docoreai_system_prompt_tokens'] = len(_sys_text_after) // 4
|
|
2211
|
+
except Exception:
|
|
2212
|
+
pass # non-critical — leave stashed value
|
|
2095
2213
|
# ============================================================
|
|
2096
2214
|
# STEP 4: SET max_tokens
|
|
2097
2215
|
# ============================================================
|
|
@@ -2162,8 +2280,15 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
2162
2280
|
from docore_ai.providers.bedrock_adapter import (
|
|
2163
2281
|
inject_bedrock_max_tokens,
|
|
2164
2282
|
)
|
|
2165
|
-
|
|
2166
|
-
|
|
2283
|
+
_has_tools = bool(kwargs.get('tools') or kwargs.get('functions'))
|
|
2284
|
+
if _has_tools:
|
|
2285
|
+
kwargs['_docoreai_injected_max_tokens'] = None
|
|
2286
|
+
dprint("🔧 Tools present — Bedrock max_tokens injection skipped")
|
|
2287
|
+
if _client_max:
|
|
2288
|
+
dprint(f"⚠️ Client has set max_tokens={_client_max} on a tool-call turn — may truncate tool-call JSON if response exceeds limit")
|
|
2289
|
+
else:
|
|
2290
|
+
inject_bedrock_max_tokens(args, final_max_tokens)
|
|
2291
|
+
kwargs['_docoreai_injected_max_tokens'] = int(final_max_tokens)
|
|
2167
2292
|
except Exception as _binj:
|
|
2168
2293
|
dprint(f"⚠️ Bedrock max_tokens injection failed: {_binj}")
|
|
2169
2294
|
else:
|
|
@@ -2190,11 +2315,25 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
2190
2315
|
kwargs['_docoreai_injected_max_tokens'] = int(final_max_tokens)
|
|
2191
2316
|
else:
|
|
2192
2317
|
_out_field = kwargs.get('_docoreai_output_field', '')
|
|
2193
|
-
|
|
2318
|
+
# ── Tool-call guard ──────────────────────────────────
|
|
2319
|
+
# If tools/functions are present, the model may return
|
|
2320
|
+
# tool-call JSON on this turn. A hard max_tokens cap
|
|
2321
|
+
# can truncate the JSON mid-structure → parse error in
|
|
2322
|
+
# the tool runner → crashed agentic loop.
|
|
2323
|
+
# Soft-limit system prompt (Step 3) is still active and
|
|
2324
|
+
# safe — the model ignores word-count hints for JSON output.
|
|
2325
|
+
_has_tools = bool(kwargs.get('tools') or kwargs.get('functions'))
|
|
2326
|
+
if _has_tools:
|
|
2327
|
+
kwargs['_docoreai_injected_max_tokens'] = None
|
|
2328
|
+
dprint("🔧 Tools present — max_tokens injection skipped to protect tool-call JSON")
|
|
2329
|
+
if _client_max:
|
|
2330
|
+
dprint(f"⚠️ Client has set max_tokens={_client_max} on a tool-call turn — may truncate tool-call JSON if response exceeds limit")
|
|
2331
|
+
elif _out_field == 'output_text':
|
|
2194
2332
|
kwargs['max_output_tokens'] = int(final_max_tokens)
|
|
2333
|
+
kwargs['_docoreai_injected_max_tokens'] = int(final_max_tokens)
|
|
2195
2334
|
else:
|
|
2196
2335
|
kwargs[_param] = int(final_max_tokens)
|
|
2197
|
-
|
|
2336
|
+
kwargs['_docoreai_injected_max_tokens'] = int(final_max_tokens)
|
|
2198
2337
|
except Exception as _inj_err:
|
|
2199
2338
|
dprint(f"⚠️ Token ceiling injection failed — call proceeds ungoverned: {_inj_err}")
|
|
2200
2339
|
except Exception as prediction_error:
|
|
@@ -2229,6 +2368,9 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
2229
2368
|
final_kwargs.pop("_docoreai_system_prompt_tokens", None)
|
|
2230
2369
|
final_kwargs.pop("_docoreai_tools_prompt_tokens", None)
|
|
2231
2370
|
final_kwargs.pop("_docoreai_tools_prompt_length", None)
|
|
2371
|
+
final_kwargs.pop("_docoreai_tool_content_original", None)
|
|
2372
|
+
final_kwargs.pop("_docoreai_tool_content_after_comp", None)
|
|
2373
|
+
final_kwargs.pop("_docoreai_tools_prompt_length_after_prune", None)
|
|
2232
2374
|
final_kwargs.pop("_docoreai_client_max_tokens", None)
|
|
2233
2375
|
final_kwargs.pop('_docoreai_injected_max_tokens', None)
|
|
2234
2376
|
final_kwargs.pop('_docoreai_governance_max_tokens', None)
|
|
@@ -2275,7 +2417,77 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
2275
2417
|
else:
|
|
2276
2418
|
dprint("🔧 No tools in final API request")
|
|
2277
2419
|
|
|
2420
|
+
# ── Prompt Compression — agentic tool results in live mode only────────────────
|
|
2421
|
+
kwargs['_docoreai_tool_content_original'] = 0
|
|
2422
|
+
kwargs['_docoreai_tool_content_after_comp'] = 0
|
|
2423
|
+
kwargs['_docoreai_tools_prompt_length_after_prune'] = kwargs.get('_docoreai_tools_prompt_length', 0)
|
|
2424
|
+
_original_tools = final_kwargs.get('tools')
|
|
2425
|
+
_pruned = None # Track HIT/MISS — guards post-call update_cache
|
|
2426
|
+
|
|
2427
|
+
if not in_learning_window:
|
|
2428
|
+
# ── Prompt Compression — agentic tool results ────────────
|
|
2429
|
+
if final_kwargs.get('messages') and final_kwargs.get('tools'):
|
|
2430
|
+
try:
|
|
2431
|
+
from docore_ai.instrumentation.agentic_cost.prompt_compression import compress_agentic_tool_results
|
|
2432
|
+
# Gate: skip log noise entirely when compression is disabled
|
|
2433
|
+
try:
|
|
2434
|
+
from docore_ai.config.db import get_phase2_config as _gpc_pc
|
|
2435
|
+
_pc_enabled = _gpc_pc('prompt_compression', 'enabled', False)
|
|
2436
|
+
except Exception:
|
|
2437
|
+
_pc_enabled = False
|
|
2438
|
+
|
|
2439
|
+
if _pc_enabled:
|
|
2440
|
+
_msgs_before = final_kwargs['messages']
|
|
2441
|
+
_chars_before = sum(len(str(m.get('content', ''))) for m in _msgs_before if isinstance(m.get('content'), str))
|
|
2442
|
+
dprint(f"🗜️ [Compression] Before: {_chars_before} chars across {len(_msgs_before)} messages")
|
|
2443
|
+
|
|
2444
|
+
final_kwargs['messages'], _pc_stats = compress_agentic_tool_results(final_kwargs['messages'])
|
|
2445
|
+
_chars_after = _pc_stats.get('chars_after', 0)
|
|
2446
|
+
_chars_orig = _pc_stats.get('chars_original', 0)
|
|
2447
|
+
dprint(f"🗜️ [Compression] After: {_chars_after} chars (saved {_chars_orig - _chars_after}, ratio {round(_chars_after/_chars_orig, 3) if _chars_orig else 'N/A'})")
|
|
2448
|
+
|
|
2449
|
+
kwargs['_docoreai_tool_content_original'] = _pc_stats.get('chars_original', 0)
|
|
2450
|
+
kwargs['_docoreai_tool_content_after_comp'] = _pc_stats.get('chars_after', 0)
|
|
2451
|
+
except Exception as _pc_err:
|
|
2452
|
+
dprint(f"⚠️ Prompt compression failed (non-fatal): {_pc_err}")
|
|
2453
|
+
|
|
2454
|
+
# ── Tool Pruning Cache — check before LLM call ──────────
|
|
2455
|
+
if _original_tools:
|
|
2456
|
+
dprint(f"✂️ [Pruning] Before: {len(_original_tools)} tools, {len(_json_before := __import__('json').dumps(_original_tools))} bytes")
|
|
2457
|
+
try:
|
|
2458
|
+
from docore_ai.instrumentation.agentic_cost.tool_pruning_cache import check_cache
|
|
2459
|
+
_pruned = check_cache(system_prompt_text, prompt_text, _original_tools, actual_provider)
|
|
2460
|
+
if _pruned is not None:
|
|
2461
|
+
final_kwargs['tools'] = _pruned
|
|
2462
|
+
import json as _json_prune
|
|
2463
|
+
_pruned_json_str = _json_prune.dumps(_pruned)
|
|
2464
|
+
kwargs['_docoreai_tools_prompt_length_after_prune'] = len(_pruned_json_str)
|
|
2465
|
+
# Update token count to reflect pruned set (mirrors Fix 1 logic)
|
|
2466
|
+
try:
|
|
2467
|
+
from docore_ai.telemetry.token_utils import _resolve_tokenizer
|
|
2468
|
+
_tok_mode, _tok_enc, _ = _resolve_tokenizer(current_model)
|
|
2469
|
+
if _tok_mode == 'tiktoken' and _tok_enc is not None:
|
|
2470
|
+
kwargs['_docoreai_tools_prompt_tokens'] = len(_tok_enc.encode(_pruned_json_str))
|
|
2471
|
+
else:
|
|
2472
|
+
kwargs['_docoreai_tools_prompt_tokens'] = len(_pruned_json_str) // 4
|
|
2473
|
+
except Exception:
|
|
2474
|
+
kwargs['_docoreai_tools_prompt_tokens'] = len(_pruned_json_str) // 4
|
|
2475
|
+
dprint(f"✂️ [Pruning] After: {len(_pruned)} tools, {kwargs['_docoreai_tools_prompt_length_after_prune']} bytes (removed {len(_original_tools) - len(_pruned)} tools)")
|
|
2476
|
+
else:
|
|
2477
|
+
dprint(f"✂️ [Pruning] Cache miss — all {len(_original_tools)} tools kept")
|
|
2478
|
+
except Exception as _tpc_err:
|
|
2479
|
+
dprint(f"⚠️ Tool pruning cache check failed (non-fatal): {_tpc_err}")
|
|
2480
|
+
else:
|
|
2481
|
+
dprint("📚 Training mode: prompt compression and tool pruning skipped")
|
|
2482
|
+
|
|
2278
2483
|
response = wrapped(*args, **final_kwargs)
|
|
2484
|
+
# ── Tool Pruning Cache — update after LLM response (Live only) ──
|
|
2485
|
+
if not in_learning_window and _original_tools and _pruned is None:
|
|
2486
|
+
try:
|
|
2487
|
+
from docore_ai.instrumentation.agentic_cost.tool_pruning_cache import update_cache
|
|
2488
|
+
update_cache(system_prompt_text, prompt_text, _original_tools, response, actual_provider)
|
|
2489
|
+
except Exception as _tpu_err:
|
|
2490
|
+
dprint(f"⚠️ Tool pruning cache update failed (non-fatal): {_tpu_err}")
|
|
2279
2491
|
dprint(f"🔍 Response type: {type(response).__name__} | has_usage={hasattr(response, 'usage')} | stream_in_kwargs={final_kwargs.get('stream', False)}")
|
|
2280
2492
|
# LangChain/OpenAI tool-calling path may return LegacyAPIResponse.
|
|
2281
2493
|
# Normalize it before DoCoreAI reads usage/text.
|
|
@@ -2443,6 +2655,30 @@ def make_llm_wrapper(provider, config_provider=None, input_field='messages', out
|
|
|
2443
2655
|
#response = wrapped(*args, **final_kwargs)
|
|
2444
2656
|
#kwargs["execution_time"] = round(time.time() - start, 2)
|
|
2445
2657
|
|
|
2658
|
+
# ====================================================
|
|
2659
|
+
# 🧠 REASONING CAPTURE (reasoning-capable models only)
|
|
2660
|
+
# ====================================================
|
|
2661
|
+
try:
|
|
2662
|
+
_reasoning_model = kwargs.get('model') or current_model
|
|
2663
|
+
dprint(f"🧠 [reasoning gate] model={_reasoning_model} extractor={reasoning_extractor} capable={_is_reasoning_capable(_reasoning_model)}")
|
|
2664
|
+
|
|
2665
|
+
if reasoning_extractor and _is_reasoning_capable(_reasoning_model):
|
|
2666
|
+
from docore_ai.reasoning_capture import extract_reasoning
|
|
2667
|
+
_reasoning_result = extract_reasoning(
|
|
2668
|
+
reasoning_extractor, kwargs, _telemetry_response
|
|
2669
|
+
)
|
|
2670
|
+
dprint(f"🧠 [reasoning result] {_reasoning_result}")
|
|
2671
|
+
kwargs['_docoreai_reasoning_effort'] = _reasoning_result.get('effort')
|
|
2672
|
+
kwargs['_docoreai_reasoning_tokens'] = _reasoning_result.get('reasoning_tokens')
|
|
2673
|
+
else:
|
|
2674
|
+
dprint(f"🧠 [reasoning skip] extractor={reasoning_extractor} capable={_is_reasoning_capable(_reasoning_model)}")
|
|
2675
|
+
kwargs['_docoreai_reasoning_effort'] = None
|
|
2676
|
+
kwargs['_docoreai_reasoning_tokens'] = None
|
|
2677
|
+
except Exception as _re:
|
|
2678
|
+
dprint(f"⚠️ Reasoning capture failed (non-fatal): {_re}")
|
|
2679
|
+
kwargs['_docoreai_reasoning_effort'] = None
|
|
2680
|
+
kwargs['_docoreai_reasoning_tokens'] = None
|
|
2681
|
+
|
|
2446
2682
|
# ====================================================
|
|
2447
2683
|
# 🛡️ GenAI Runtime Governance – REGISTER FINGERPRINTS
|
|
2448
2684
|
# ====================================================
|
|
@@ -2696,6 +2932,10 @@ def log_failed_telemetry(args, kwargs, provider, error, in_learning_window=True)
|
|
|
2696
2932
|
client_prompt_id=kwargs.get('_docoreai_client_prompt_id'),
|
|
2697
2933
|
soft_limit_tokens_applied=kwargs.get('_docoreai_soft_limit_tokens_applied',0),
|
|
2698
2934
|
soft_limit_target_pct_reduced=float(get_phase2_config('soft_limit', 'target_reduction_pct', 0.0)),
|
|
2935
|
+
tool_content_original = kwargs.get('_docoreai_tool_content_original', 0),
|
|
2936
|
+
tool_content_after_comp = kwargs.get('_docoreai_tool_content_after_comp', 0),
|
|
2937
|
+
tools_prompt_length_after_prune= kwargs.get('_docoreai_tools_prompt_length_after_prune', 0),
|
|
2938
|
+
|
|
2699
2939
|
)
|
|
2700
2940
|
|
|
2701
2941
|
# ✅ Add error details to payload
|
|
@@ -2783,7 +3023,8 @@ def _load_patch_targets() -> list:
|
|
|
2783
3023
|
"provider, "
|
|
2784
3024
|
"response_format, "
|
|
2785
3025
|
"COALESCE(input_field, 'messages'), "
|
|
2786
|
-
"COALESCE(output_field, 'choices[0].message.content') "
|
|
3026
|
+
"COALESCE(output_field, 'choices[0].message.content'), "
|
|
3027
|
+
"reasoning_extractor "
|
|
2787
3028
|
"FROM wp_docoreai_patch_targets "
|
|
2788
3029
|
"WHERE enabled = 1 "
|
|
2789
3030
|
"ORDER BY id ASC"
|
|
@@ -2799,7 +3040,8 @@ def _load_patch_targets() -> list:
|
|
|
2799
3040
|
r[2],
|
|
2800
3041
|
r[3],
|
|
2801
3042
|
r[4] if len(r) > 4 else 'messages',
|
|
2802
|
-
r[5] if len(r) > 5 else 'choices[0].message.content'
|
|
3043
|
+
r[5] if len(r) > 5 else 'choices[0].message.content',
|
|
3044
|
+
r[6] if len(r) > 6 else None
|
|
2803
3045
|
)
|
|
2804
3046
|
for r in rows
|
|
2805
3047
|
]
|
|
@@ -2891,7 +3133,11 @@ def apply_patches():
|
|
|
2891
3133
|
|
|
2892
3134
|
for target in active_targets:
|
|
2893
3135
|
|
|
2894
|
-
#
|
|
3136
|
+
# 7-tuple from DB:
|
|
3137
|
+
# (module_path, class_method, provider,
|
|
3138
|
+
# response_format, input_field, output_field, reasoning_extractor)
|
|
3139
|
+
#
|
|
3140
|
+
# 6-tuple from DB (legacy):
|
|
2895
3141
|
# (module_path, class_method, provider,
|
|
2896
3142
|
# response_format, input_field, output_field)
|
|
2897
3143
|
#
|
|
@@ -2901,7 +3147,20 @@ def apply_patches():
|
|
|
2901
3147
|
# 3-tuple compatibility:
|
|
2902
3148
|
# (module_path, class_method, provider)
|
|
2903
3149
|
|
|
2904
|
-
|
|
3150
|
+
reasoning_extractor = None # default
|
|
3151
|
+
|
|
3152
|
+
if len(target) == 7:
|
|
3153
|
+
(
|
|
3154
|
+
module_path,
|
|
3155
|
+
attr_path,
|
|
3156
|
+
provider,
|
|
3157
|
+
response_format,
|
|
3158
|
+
input_field,
|
|
3159
|
+
output_field,
|
|
3160
|
+
reasoning_extractor
|
|
3161
|
+
) = target
|
|
3162
|
+
|
|
3163
|
+
elif len(target) == 6:
|
|
2905
3164
|
(
|
|
2906
3165
|
module_path,
|
|
2907
3166
|
attr_path,
|
|
@@ -2961,7 +3220,8 @@ def apply_patches():
|
|
|
2961
3220
|
make_llm_wrapper(
|
|
2962
3221
|
provider,
|
|
2963
3222
|
input_field=input_field,
|
|
2964
|
-
output_field=output_field
|
|
3223
|
+
output_field=output_field,
|
|
3224
|
+
reasoning_extractor=reasoning_extractor
|
|
2965
3225
|
)
|
|
2966
3226
|
)
|
|
2967
3227
|
|
|
@@ -281,6 +281,7 @@ def start():
|
|
|
281
281
|
|
|
282
282
|
if not _check_membership():
|
|
283
283
|
dprint("🛑 DoCoreAI cannot start — membership invalid.")
|
|
284
|
+
print("🛑 DoCoreAI stopped — membership invalid. Check the log file for details.", flush=True)
|
|
284
285
|
return
|
|
285
286
|
|
|
286
287
|
# Start auto-retraining monitor (Phase 2.2)
|
|
@@ -386,7 +387,37 @@ def start():
|
|
|
386
387
|
# dprint("📚 Learning phase active - budget not required yet")
|
|
387
388
|
# dprint(f" {get_prediction_count()}/100 predictions collected")
|
|
388
389
|
#dprint("✅ All systems ready")
|
|
389
|
-
|
|
390
|
+
# Pre-warm tool pruning embedder (downloads model once, cached forever after)
|
|
391
|
+
try:
|
|
392
|
+
if get_phase2_config('tool_pruning', 'enabled', True):
|
|
393
|
+
from docore_ai.instrumentation.agentic_cost.embedder import encode
|
|
394
|
+
encode('')
|
|
395
|
+
dprint("✅ Tool pruning embedder pre-loaded")
|
|
396
|
+
except Exception as e:
|
|
397
|
+
dprint(f"⚠️ Tool pruning embedder warm-up skipped: {e}")
|
|
398
|
+
|
|
399
|
+
# Pre-warm LLMLingua-2 prompt compressor in a background thread
|
|
400
|
+
# Model cold load takes 60–190 s — must NOT block the start command
|
|
401
|
+
try:
|
|
402
|
+
_pc_en = str(get_phase2_config('prompt_compression', 'enabled', 'false')).lower() == 'true'
|
|
403
|
+
_pc_ag = str(get_phase2_config('prompt_compression', 'agentic_enabled', 'false')).lower() == 'true'
|
|
404
|
+
_pc_rt = get_phase2_config('prompt_compression', 'agentic_rate', '0.9')
|
|
405
|
+
_pc_status = "ENABLED ✅" if (_pc_en and _pc_ag) else "DISABLED ⛔"
|
|
406
|
+
dprint(f"🗜️ Prompt compression: {_pc_status} | enabled={_pc_en}, agentic={_pc_ag}, rate={_pc_rt}")
|
|
407
|
+
if _pc_en and _pc_ag:
|
|
408
|
+
import threading as _threading
|
|
409
|
+
def _warm_compressor():
|
|
410
|
+
try:
|
|
411
|
+
from docore_ai.instrumentation.agentic_cost.prompt_compression import preload_compressor
|
|
412
|
+
preload_compressor()
|
|
413
|
+
except Exception as _wc_err:
|
|
414
|
+
dprint(f"⚠️ Prompt compressor warm-up failed: {_wc_err}")
|
|
415
|
+
_t = _threading.Thread(target=_warm_compressor, daemon=True, name="compressor-preload")
|
|
416
|
+
_t.start()
|
|
417
|
+
dprint("🗜️ Prompt compressor pre-load started (background thread)")
|
|
418
|
+
except Exception as e:
|
|
419
|
+
dprint(f"⚠️ Prompt compressor warm-up skipped: {e}")
|
|
420
|
+
show_start_banner()
|
|
390
421
|
|
|
391
422
|
|
|
392
423
|
# 🔴 FINAL STEP: Trigger autopatch via .pth
|
|
@@ -1016,6 +1047,7 @@ def _check_membership() -> bool:
|
|
|
1016
1047
|
|
|
1017
1048
|
if state == "expired":
|
|
1018
1049
|
dprint(f"❌ Membership expired — please renew at docoreai.com")
|
|
1050
|
+
print("❌ DoCoreAI cannot start — membership expired. Please renew at docoreai.com", flush=True)
|
|
1019
1051
|
return False
|
|
1020
1052
|
# unknown/other states — allow through (endpoint may be returning partial data)
|
|
1021
1053
|
if state != "active":
|