Files
ss-tools/backend/src/plugins/llm_analysis/plugin.py

470 lines
22 KiB
Python

# [DEF:backend/src/plugins/llm_analysis/plugin.py:Module]
# @TIER: STANDARD
# @SEMANTICS: plugin, llm, analysis, documentation
# @PURPOSE: Implements DashboardValidationPlugin and DocumentationPlugin.
# @LAYER: Domain
# @RELATION: INHERITS -> backend.src.core.plugin_base.PluginBase
# @RELATION: CALLS -> backend.src.plugins.llm_analysis.service.ScreenshotService
# @RELATION: CALLS -> backend.src.plugins.llm_analysis.service.LLMClient
# @RELATION: CALLS -> backend.src.services.llm_provider.LLMProviderService
# @RELATION: USES -> TaskContext
# @INVARIANT: All LLM interactions must be executed as asynchronous tasks.
from typing import Dict, Any, Optional
import os
import json
from datetime import datetime, timedelta
from ...core.plugin_base import PluginBase
from ...core.logger import belief_scope, logger
from ...core.database import SessionLocal
from ...services.llm_provider import LLMProviderService
from ...core.superset_client import SupersetClient
from .service import ScreenshotService, LLMClient
from .models import LLMProviderType, ValidationStatus, ValidationResult, DetectedIssue
from ...models.llm import ValidationRecord
from ...core.task_manager.context import TaskContext
from ...services.llm_prompt_templates import (
DEFAULT_LLM_PROMPTS,
is_multimodal_model,
normalize_llm_settings,
render_prompt,
)
# [DEF:_is_masked_or_invalid_api_key:Function]
# @PURPOSE: Guards against placeholder or malformed API keys in runtime.
# @PRE: value may be None.
# @POST: Returns True when value cannot be used for authenticated provider calls.
def _is_masked_or_invalid_api_key(value: Optional[str]) -> bool:
key = (value or "").strip()
if not key:
return True
if key in {"********", "EMPTY_OR_NONE"}:
return True
# Most provider tokens are significantly longer; short values are almost always placeholders.
return len(key) < 16
# [/DEF:_is_masked_or_invalid_api_key:Function]
# [DEF:_json_safe_value:Function]
# @PURPOSE: Recursively normalize payload values for JSON serialization.
# @PRE: value may be nested dict/list with datetime values.
# @POST: datetime values are converted to ISO strings.
def _json_safe_value(value: Any):
if isinstance(value, datetime):
return value.isoformat()
if isinstance(value, dict):
return {k: _json_safe_value(v) for k, v in value.items()}
if isinstance(value, list):
return [_json_safe_value(v) for v in value]
return value
# [/DEF:_json_safe_value:Function]
# [DEF:DashboardValidationPlugin:Class]
# @PURPOSE: Plugin for automated dashboard health analysis using LLMs.
# @RELATION: IMPLEMENTS -> backend.src.core.plugin_base.PluginBase
class DashboardValidationPlugin(PluginBase):
@property
def id(self) -> str:
return "llm_dashboard_validation"
@property
def name(self) -> str:
return "Dashboard LLM Validation"
@property
def description(self) -> str:
return "Automated dashboard health analysis using multimodal LLMs."
@property
def version(self) -> str:
return "1.0.0"
def get_schema(self) -> Dict[str, Any]:
return {
"type": "object",
"properties": {
"dashboard_id": {"type": "string", "title": "Dashboard ID"},
"environment_id": {"type": "string", "title": "Environment ID"},
"provider_id": {"type": "string", "title": "LLM Provider ID"}
},
"required": ["dashboard_id", "environment_id", "provider_id"]
}
# [DEF:DashboardValidationPlugin.execute:Function]
# @PURPOSE: Executes the dashboard validation task with TaskContext support.
# @PARAM: params (Dict[str, Any]) - Validation parameters.
# @PARAM: context (Optional[TaskContext]) - Task context for logging with source attribution.
# @PRE: params contains dashboard_id, environment_id, and provider_id.
# @POST: Returns a dictionary with validation results and persists them to the database.
# @SIDE_EFFECT: Captures a screenshot, calls LLM API, and writes to the database.
async def execute(self, params: Dict[str, Any], context: Optional[TaskContext] = None):
with belief_scope("execute", f"plugin_id={self.id}"):
validation_started_at = datetime.utcnow()
# Use TaskContext logger if available, otherwise fall back to app logger
log = context.logger if context else logger
# Create sub-loggers for different components
llm_log = log.with_source("llm") if context else log
screenshot_log = log.with_source("screenshot") if context else log
superset_log = log.with_source("superset_api") if context else log
log.info(f"Executing {self.name} with params: {params}")
dashboard_id_raw = params.get("dashboard_id")
dashboard_id = str(dashboard_id_raw) if dashboard_id_raw is not None else None
env_id = params.get("environment_id")
provider_id = params.get("provider_id")
db = SessionLocal()
try:
# 1. Get Environment
from ...dependencies import get_config_manager
config_mgr = get_config_manager()
env = config_mgr.get_environment(env_id)
if not env:
log.error(f"Environment {env_id} not found")
raise ValueError(f"Environment {env_id} not found")
# 2. Get LLM Provider
llm_service = LLMProviderService(db)
db_provider = llm_service.get_provider(provider_id)
if not db_provider:
log.error(f"LLM Provider {provider_id} not found")
raise ValueError(f"LLM Provider {provider_id} not found")
llm_log.debug("Retrieved provider config:")
llm_log.debug(f" Provider ID: {db_provider.id}")
llm_log.debug(f" Provider Name: {db_provider.name}")
llm_log.debug(f" Provider Type: {db_provider.provider_type}")
llm_log.debug(f" Base URL: {db_provider.base_url}")
llm_log.debug(f" Default Model: {db_provider.default_model}")
llm_log.debug(f" Is Active: {db_provider.is_active}")
if not is_multimodal_model(db_provider.default_model, db_provider.provider_type):
raise ValueError(
"Dashboard validation requires a multimodal model (image input support)."
)
api_key = llm_service.get_decrypted_api_key(provider_id)
llm_log.debug(f"API Key decrypted (first 8 chars): {api_key[:8] if api_key and len(api_key) > 8 else 'EMPTY_OR_NONE'}...")
# Check if API key was successfully decrypted
if _is_masked_or_invalid_api_key(api_key):
raise ValueError(
f"Invalid API key for provider {provider_id}. "
"Please open LLM provider settings and save a real API key (not masked placeholder)."
)
# 3. Capture Screenshot
screenshot_service = ScreenshotService(env)
storage_root = config_mgr.get_config().settings.storage.root_path
screenshots_dir = os.path.join(storage_root, "screenshots")
os.makedirs(screenshots_dir, exist_ok=True)
filename = f"{dashboard_id}_{datetime.now().strftime('%Y%m%d_%H%M%S')}.png"
screenshot_path = os.path.join(screenshots_dir, filename)
screenshot_started_at = datetime.utcnow()
screenshot_log.info(f"Capturing screenshot for dashboard {dashboard_id}")
await screenshot_service.capture_dashboard(dashboard_id, screenshot_path)
screenshot_log.debug(f"Screenshot saved to: {screenshot_path}")
screenshot_finished_at = datetime.utcnow()
# 4. Fetch Logs (from Environment /api/v1/log/)
logs = []
logs_fetch_started_at = datetime.utcnow()
try:
client = SupersetClient(env)
# Calculate time window (last 24 hours)
start_time = (datetime.now() - timedelta(hours=24)).isoformat()
# Construct filter for logs
# Note: We filter by dashboard_id matching the object
query_params = {
"filters": [
{"col": "dashboard_id", "opr": "eq", "value": dashboard_id},
{"col": "dttm", "opr": "gt", "value": start_time}
],
"order_column": "dttm",
"order_direction": "desc",
"page": 0,
"page_size": 100
}
superset_log.debug(f"Fetching logs for dashboard {dashboard_id}")
response = client.network.request(
method="GET",
endpoint="/log/",
params={"q": json.dumps(query_params)}
)
if isinstance(response, dict) and "result" in response:
for item in response["result"]:
action = item.get("action", "unknown")
dttm = item.get("dttm", "")
details = item.get("json", "")
logs.append(f"[{dttm}] {action}: {details}")
if not logs:
logs = ["No recent logs found for this dashboard."]
superset_log.debug("No recent logs found for this dashboard")
except Exception as e:
superset_log.warning(f"Failed to fetch logs from environment: {e}")
logs = [f"Error fetching remote logs: {str(e)}"]
logs_fetch_finished_at = datetime.utcnow()
# 5. Analyze with LLM
llm_client = LLMClient(
provider_type=LLMProviderType(db_provider.provider_type),
api_key=api_key,
base_url=db_provider.base_url,
default_model=db_provider.default_model
)
llm_log.info(f"Analyzing dashboard {dashboard_id} with LLM")
llm_settings = normalize_llm_settings(config_mgr.get_config().settings.llm)
dashboard_prompt = llm_settings["prompts"].get(
"dashboard_validation_prompt",
DEFAULT_LLM_PROMPTS["dashboard_validation_prompt"],
)
llm_call_started_at = datetime.utcnow()
analysis = await llm_client.analyze_dashboard(
screenshot_path,
logs,
prompt_template=dashboard_prompt,
)
llm_call_finished_at = datetime.utcnow()
# Log analysis summary to task logs for better visibility
llm_log.info(f"[ANALYSIS_SUMMARY] Status: {analysis['status']}")
llm_log.info(f"[ANALYSIS_SUMMARY] Summary: {analysis['summary']}")
if analysis.get("issues"):
for i, issue in enumerate(analysis["issues"]):
llm_log.info(f"[ANALYSIS_ISSUE][{i+1}] {issue.get('severity')}: {issue.get('message')} (Location: {issue.get('location', 'N/A')})")
# 6. Persist Result
validation_result = ValidationResult(
dashboard_id=dashboard_id,
status=ValidationStatus(analysis["status"]),
summary=analysis["summary"],
issues=[DetectedIssue(**issue) for issue in analysis["issues"]],
screenshot_path=screenshot_path,
raw_response=str(analysis)
)
validation_finished_at = datetime.utcnow()
result_payload = _json_safe_value(validation_result.dict())
result_payload["screenshot_paths"] = [screenshot_path]
result_payload["logs_sent_to_llm"] = logs
result_payload["logs_sent_count"] = len(logs)
result_payload["prompt_template"] = dashboard_prompt
result_payload["provider"] = {
"id": db_provider.id,
"name": db_provider.name,
"type": db_provider.provider_type,
"base_url": db_provider.base_url,
"model": db_provider.default_model,
}
result_payload["environment_id"] = env_id
result_payload["timings"] = {
"validation_started_at": validation_started_at.isoformat(),
"validation_finished_at": validation_finished_at.isoformat(),
"validation_duration_ms": int((validation_finished_at - validation_started_at).total_seconds() * 1000),
"screenshot_started_at": screenshot_started_at.isoformat(),
"screenshot_finished_at": screenshot_finished_at.isoformat(),
"screenshot_duration_ms": int((screenshot_finished_at - screenshot_started_at).total_seconds() * 1000),
"logs_fetch_started_at": logs_fetch_started_at.isoformat(),
"logs_fetch_finished_at": logs_fetch_finished_at.isoformat(),
"logs_fetch_duration_ms": int((logs_fetch_finished_at - logs_fetch_started_at).total_seconds() * 1000),
"llm_call_started_at": llm_call_started_at.isoformat(),
"llm_call_finished_at": llm_call_finished_at.isoformat(),
"llm_call_duration_ms": int((llm_call_finished_at - llm_call_started_at).total_seconds() * 1000),
}
db_record = ValidationRecord(
dashboard_id=validation_result.dashboard_id,
status=validation_result.status.value,
summary=validation_result.summary,
issues=[issue.dict() for issue in validation_result.issues],
screenshot_path=validation_result.screenshot_path,
raw_response=json.dumps(result_payload, ensure_ascii=False)
)
db.add(db_record)
db.commit()
# 7. Notification on failure (US1 / FR-015)
if validation_result.status == ValidationStatus.FAIL:
log.warning(f"Dashboard {dashboard_id} validation FAILED. Summary: {validation_result.summary}")
# Placeholder for Email/Pulse notification dispatch
# In a real implementation, we would call a NotificationService here
# with a payload containing the summary and a link to the report.
# Final log to ensure all analysis is visible in task logs
log.info(f"Validation completed for dashboard {dashboard_id}. Status: {validation_result.status.value}")
return result_payload
finally:
db.close()
# [/DEF:DashboardValidationPlugin.execute:Function]
# [/DEF:DashboardValidationPlugin:Class]
# [DEF:DocumentationPlugin:Class]
# @PURPOSE: Plugin for automated dataset documentation using LLMs.
# @RELATION: IMPLEMENTS -> backend.src.core.plugin_base.PluginBase
class DocumentationPlugin(PluginBase):
@property
def id(self) -> str:
return "llm_documentation"
@property
def name(self) -> str:
return "Dataset LLM Documentation"
@property
def description(self) -> str:
return "Automated dataset and column documentation using LLMs."
@property
def version(self) -> str:
return "1.0.0"
def get_schema(self) -> Dict[str, Any]:
return {
"type": "object",
"properties": {
"dataset_id": {"type": "string", "title": "Dataset ID"},
"environment_id": {"type": "string", "title": "Environment ID"},
"provider_id": {"type": "string", "title": "LLM Provider ID"}
},
"required": ["dataset_id", "environment_id", "provider_id"]
}
# [DEF:DocumentationPlugin.execute:Function]
# @PURPOSE: Executes the dataset documentation task with TaskContext support.
# @PARAM: params (Dict[str, Any]) - Documentation parameters.
# @PARAM: context (Optional[TaskContext]) - Task context for logging with source attribution.
# @PRE: params contains dataset_id, environment_id, and provider_id.
# @POST: Returns generated documentation and updates the dataset in Superset.
# @SIDE_EFFECT: Calls LLM API and updates dataset metadata in Superset.
async def execute(self, params: Dict[str, Any], context: Optional[TaskContext] = None):
with belief_scope("execute", f"plugin_id={self.id}"):
# Use TaskContext logger if available, otherwise fall back to app logger
log = context.logger if context else logger
# Create sub-loggers for different components
llm_log = log.with_source("llm") if context else log
superset_log = log.with_source("superset_api") if context else log
log.info(f"Executing {self.name} with params: {params}")
dataset_id = params.get("dataset_id")
env_id = params.get("environment_id")
provider_id = params.get("provider_id")
db = SessionLocal()
try:
# 1. Get Environment
from ...dependencies import get_config_manager
config_mgr = get_config_manager()
env = config_mgr.get_environment(env_id)
if not env:
log.error(f"Environment {env_id} not found")
raise ValueError(f"Environment {env_id} not found")
# 2. Get LLM Provider
llm_service = LLMProviderService(db)
db_provider = llm_service.get_provider(provider_id)
if not db_provider:
log.error(f"LLM Provider {provider_id} not found")
raise ValueError(f"LLM Provider {provider_id} not found")
llm_log.debug("Retrieved provider config:")
llm_log.debug(f" Provider ID: {db_provider.id}")
llm_log.debug(f" Provider Name: {db_provider.name}")
llm_log.debug(f" Provider Type: {db_provider.provider_type}")
llm_log.debug(f" Base URL: {db_provider.base_url}")
llm_log.debug(f" Default Model: {db_provider.default_model}")
api_key = llm_service.get_decrypted_api_key(provider_id)
llm_log.debug(f"API Key decrypted (first 8 chars): {api_key[:8] if api_key and len(api_key) > 8 else 'EMPTY_OR_NONE'}...")
# Check if API key was successfully decrypted
if _is_masked_or_invalid_api_key(api_key):
raise ValueError(
f"Invalid API key for provider {provider_id}. "
"Please open LLM provider settings and save a real API key (not masked placeholder)."
)
# 3. Fetch Metadata (US2 / T024)
from ...core.superset_client import SupersetClient
client = SupersetClient(env)
superset_log.debug(f"Fetching dataset {dataset_id}")
dataset = client.get_dataset(int(dataset_id))
# Extract columns and existing descriptions
columns_data = []
for col in dataset.get("columns", []):
columns_data.append({
"name": col.get("column_name"),
"type": col.get("type"),
"description": col.get("description")
})
superset_log.debug(f"Extracted {len(columns_data)} columns from dataset")
# 4. Construct Prompt & Analyze (US2 / T025)
llm_client = LLMClient(
provider_type=LLMProviderType(db_provider.provider_type),
api_key=api_key,
base_url=db_provider.base_url,
default_model=db_provider.default_model
)
llm_settings = normalize_llm_settings(config_mgr.get_config().settings.llm)
documentation_prompt = llm_settings["prompts"].get(
"documentation_prompt",
DEFAULT_LLM_PROMPTS["documentation_prompt"],
)
prompt = render_prompt(
documentation_prompt,
{
"dataset_name": dataset.get("table_name") or "",
"columns_json": json.dumps(columns_data, ensure_ascii=False),
},
)
# Using a generic chat completion for text-only US2
llm_log.info(f"Generating documentation for dataset {dataset_id}")
doc_result = await llm_client.get_json_completion([{"role": "user", "content": prompt}])
# 5. Update Metadata (US2 / T026)
update_payload = {
"description": doc_result["dataset_description"],
"columns": []
}
# Map generated descriptions back to column IDs
for col_doc in doc_result["column_descriptions"]:
for col in dataset.get("columns", []):
if col.get("column_name") == col_doc["name"]:
update_payload["columns"].append({
"id": col.get("id"),
"description": col_doc["description"]
})
superset_log.info(f"Updating dataset {dataset_id} with generated documentation")
client.update_dataset(int(dataset_id), update_payload)
log.info(f"Documentation completed for dataset {dataset_id}")
return doc_result
finally:
db.close()
# [/DEF:DocumentationPlugin.execute:Function]
# [/DEF:DocumentationPlugin:Class]
# [/DEF:backend/src/plugins/llm_analysis/plugin.py:Module]