Files

154 lines
5.0 KiB
Python

"""LLM analysis client for GuardDog findings.
Supports any OpenAI-compatible API endpoint with configurable model.
"""
import asyncio
import json
import httpx
from ..config import config
from ..constants import LLM_ANALYSIS_SYSTEM_PROMPT, LLM_DEFAULT_TEMPERATURE, LLM_RESPONSE_FORMAT
from ..logging_setup import log
_llm_semaphore = asyncio.Semaphore(config.llm_max_concurrent)
_REPORT_DEFAULTS = {
"verdict": "unknown",
"summary": "No summary provided",
"analysis": "No analysis provided",
"severity_rating": "unknown",
}
def _validate_report(report: dict) -> dict:
result = dict(report)
for field, default in _REPORT_DEFAULTS.items():
if not result.get(field):
result[field] = default
if result["verdict"] not in ("safe", "suspicious", "malicious", "unknown"):
result["verdict"] = "unknown"
return result
def _build_user_message(finding: dict) -> str:
"""Build a concise prompt from a finding's data."""
rule = finding.get("rule", "unknown")
severity = finding.get("severity", "unknown")
message = finding.get("message", "")
location = finding.get("location", "")
code = finding.get("code", "")
prompt = f"Rule: {rule}\nSeverity: {severity}\nMessage: {message}\n"
if location:
prompt += f"Location: {location}\n"
if code:
prompt += f"Code snippet:\n```\n{code}\n```\n"
prompt += (
"\nAnalyse this finding and return JSON with keys: "
"verdict, summary, analysis, severity_rating."
)
return prompt
async def _attempt_llm_call(finding_data: dict) -> dict | None:
"""Single attempt to call LLM and parse response."""
url = f"{config.llm_api_base.rstrip('/')}/chat/completions"
headers = {
"Authorization": f"Bearer {config.llm_api_key}",
"Content-Type": "application/json",
}
payload = {
"model": config.llm_model,
"messages": [
{"role": "system", "content": LLM_ANALYSIS_SYSTEM_PROMPT},
{"role": "user", "content": _build_user_message(finding_data)},
],
"temperature": LLM_DEFAULT_TEMPERATURE,
"response_format": {"type": LLM_RESPONSE_FORMAT},
}
try:
async with _llm_semaphore:
async with httpx.AsyncClient(timeout=config.llm_timeout, headers=headers) as client:
resp = await client.post(url, json=payload)
resp.raise_for_status()
body = resp.json()
except httpx.TimeoutException:
log.error(
"LLM analysis timed out after %ds for rule=%s",
config.llm_timeout,
finding_data.get("rule"),
)
return None
except Exception as e:
log.warning("LLM analysis failed for rule=%s: %s", finding_data.get("rule"), e)
return None
try:
choices = body.get("choices", [])
if not choices:
raise ValueError("Empty choices list")
message = choices[0].get("message", {})
content = message.get("content", "")
if not content:
raise ValueError("Empty message content")
parsed = json.loads(content)
return _validate_report(parsed)
except (ValueError, json.JSONDecodeError) as e:
raw = ""
try:
choices = body.get("choices", [])
if choices:
message = choices[0].get("message", {})
raw = message.get("content", "")
except (KeyError, IndexError):
raw = str(body)[:300]
# Some models wrap JSON in markdown code blocks
if isinstance(raw, str) and raw.strip().startswith("```"):
try:
stripped = raw.strip().strip("`").strip()
if stripped.startswith("json\n"):
stripped = stripped[5:]
return _validate_report(json.loads(stripped))
except json.JSONDecodeError:
pass
log.warning(
"LLM response parse error for rule=%s: %s — raw=%s",
finding_data.get("rule"),
e,
raw[:200] if isinstance(raw, str) else str(raw)[:200],
)
return None
async def analyze_finding(finding_data: dict, max_retries: int = 3) -> dict | None:
"""Send a finding to the LLM for security analysis with retry logic.
Returns parsed JSON dict on success, or None on failure.
"""
if not config.llm_api_key:
log.warning("LLM_API_KEY not set — skipping LLM analysis")
return None
for attempt in range(max_retries):
result = await _attempt_llm_call(finding_data)
if result is not None:
return result
if attempt < max_retries - 1:
await asyncio.sleep(2**attempt * 2) # 2s, 4s, 8s
log.info(
"Retrying LLM analysis for rule=%s (attempt %d)",
finding_data.get("rule"),
attempt + 2,
)
log.error(
"LLM analysis failed after %d attempts for rule=%s",
max_retries,
finding_data.get("rule"),
)
return None