Files
hermes-agent/tools/web_tools_extract.py

212 lines
10 KiB
Python

"""web_extract helpers: URL validation, provider resolution, cache-aware dispatch.
Order of controls (each is a gate, never skipped by a cache hit): secret-URL
refusal -> SSRF filter (in web_tools.web_extract_tool) -> provider resolution
(strict selection) -> per-URL website policy -> disk cache -> vendor call with
one-shot keyless rescue. Names are re-imported by tools/web_tools.py so
``tools.web_tools._validate_extract_urls`` etc. keep working; logs under the origin logger.
"""
import asyncio
import json
import logging
from typing import Any, Dict, List, Optional
from tools.tool_backend_helpers import selection_error, selection_exists
from tools.url_safety import normalize_url_for_request, sensitive_query_param_name
from tools.web_tools_rescue import _rescue_eligible, _rescue_extract
logger = logging.getLogger("tools.web_tools")
_NO_RESULT_ERROR = "Extract backend returned no result for this URL"
_EXTRACT_BACKENDS_HINT = "firecrawl, tavily, keenable, exa, or parallel."
_INVALID_ITEM_ERROR = (
"Invalid URL item at index {}: expected a URL string or an object with a string 'url' or 'href' field"
)
def _web_extract_url(value: Any) -> Optional[str]:
"""URL from a model-supplied extract item (str, or dict with ``url``/``href``); None if unusable.
Models sometimes forward a whole search result instead of its URL, hence the dict form. Never
stringify arbitrary objects into misleading fetch targets.
"""
if isinstance(value, dict):
value = value.get("url") or value.get("href")
return (value.strip() or None) if isinstance(value, str) else None
def _disabled_plugin_error(capability: str, disabled_key: str) -> str:
"""Error text when the configured backend's bundled plugin is disabled in config."""
vendor = disabled_key.split("/", 1)[-1]
return (
f"web.{capability}_backend is set to '{vendor}', but its plugin ('{disabled_key}') is disabled "
f"in config. Re-enable it with `hermes plugins enable {disabled_key}` "
"(or remove it from plugins.disabled)."
)
def _no_provider_error(capability: str, fallback: str) -> str:
"""Error when no provider resolved: point at a disabled bundled plugin if that is the real cause."""
from agent.web_search_registry import _disabled_web_plugin_for
disabled_key = _disabled_web_plugin_for(capability=capability)
return _disabled_plugin_error(capability, disabled_key) if disabled_key else fallback
def _strict_selection_error(capability: str, backend: str) -> str:
"""Error for a stored-but-unregistered backend: name the disabled plugin, else the bad selection.
Strict selection never silently switches to whatever the availability walk finds."""
failure = f"no registered web {capability} provider has that name"
fallback = selection_error("web", f"'{backend}'", failure)
return _no_provider_error(capability, fallback)
def _result_entry(url: str, error: Optional[str]) -> Dict[str, Any]:
return {"url": url, "title": "", "content": "", "error": error}
def _extract_error_json(error: str) -> str:
return json.dumps({"success": False, "error": error}, ensure_ascii=False)
def _refuse_all(error: str):
"""Whole-call refusal tuple for ``_validate_extract_urls`` (exfiltration prevention)."""
return None, None, None, json.dumps({"success": False, "error": error})
def _merge_in_order(
total: int, fixed: Dict[int, dict], fetch_positions: List[int], fetch_urls: List[str], results: List[dict]
) -> List[dict]:
"""Rebuild a ``total``-long result list: *fixed* entries by position, fetched *results* at
*fetch_positions* (a short provider list yields ``_NO_RESULT_ERROR`` entries for the rest)."""
merged = dict(fixed)
for pos, position in enumerate(fetch_positions):
missing = _result_entry(fetch_urls[pos], _NO_RESULT_ERROR)
merged[position] = results[pos] if pos < len(results) else missing
return [merged[i] for i in range(total)]
def _validate_extract_urls(urls: List[Any]):
"""Normalize model-supplied items and block URLs carrying secrets (percent-encoded forms are unquoted
and checked too). Returns ``(normalized_urls, normalized_indices, invalid_urls, blocked_json)``;
``blocked_json`` is a whole-call refusal (exfiltration prevention) or None."""
from agent.redact import _PREFIX_RE
from urllib.parse import unquote
normalized_urls: List[str] = []
normalized_indices: List[int] = []
invalid_urls: Dict[int, Dict[str, Any]] = {}
for index, item in enumerate(urls):
_url = _web_extract_url(item)
if _url is None:
invalid_urls[index] = _result_entry("", _INVALID_ITEM_ERROR.format(index))
continue
normalized_url = normalize_url_for_request(_url)
if any(_PREFIX_RE.search(c) for c in (_url, unquote(_url), normalized_url, unquote(normalized_url))):
return _refuse_all(
"Blocked: URL contains what appears to be an API key or token. "
"Secrets must not be sent in URLs."
)
sensitive_query_key = sensitive_query_param_name(normalized_url)
if sensitive_query_key:
return _refuse_all(
"Blocked: URL contains a credential-like query parameter "
f"({sensitive_query_key}). Web extract backends are third-party "
"readers; remove the sensitive query parameter or use a local "
"browser session when this access is explicitly required."
)
normalized_urls.append(normalized_url)
normalized_indices.append(index)
return normalized_urls, normalized_indices, invalid_urls, None
def _resolve_extract_provider(backend: str):
"""Resolve the extract provider for *backend*; returns ``(provider, error_json)``.
A registered search-only backend is a typed error (never a silent switch). An unregistered name with
a stored web selection is a strict-selection error; with no selection, fall through to the walk.
"""
from agent.web_search_registry import get_active_extract_provider, get_provider as _wsp_get_provider
provider = _wsp_get_provider(backend) if backend else None
if provider is not None and provider.supports_extract():
return provider, None
if provider is not None:
return None, _extract_error_json(
f"{provider.display_name} is a search-only backend and cannot extract URL content. "
"Set web.extract_backend to " + _EXTRACT_BACKENDS_HINT
)
if backend and selection_exists("web"):
return None, _extract_error_json(_strict_selection_error("extract", backend))
provider = get_active_extract_provider()
if provider is None:
return None, _extract_error_json(_no_provider_error(
"extract",
"No web extract provider configured. Set web.extract_backend to " + _EXTRACT_BACKENDS_HINT,
))
return provider, None
async def _dispatch_extract(provider, fetch_urls: List[str], format: Optional[str]) -> List[dict]:
"""Call ``provider.extract`` (async or sync-in-thread), with one-shot keyless rescue.
Rescue fires on a raised exception or when the WHOLE batch failed (backend outage, not per-page
problems). Rescued batches are never cached.
"""
import inspect
from tools.web_result_cache import extract_cache_put
try:
if inspect.iscoroutinefunction(provider.extract):
results = await provider.extract(fetch_urls, format=format)
else: # sync extract() runs in a thread so network I/O never blocks the loop
results = await asyncio.to_thread(provider.extract, fetch_urls, format=format)
except Exception as exc: # noqa: BLE001 — candidate for rescue
if not _rescue_eligible(provider):
raise
failed = [_result_entry(u, str(exc)) for u in fetch_urls]
return await asyncio.to_thread(_rescue_extract, provider.name, fetch_urls, failed)
if results and all(r.get("error") for r in results) and _rescue_eligible(provider):
return await asyncio.to_thread(_rescue_extract, provider.name, fetch_urls, results)
# Cache each successful fetch's full clean text (best-effort; oversized skipped).
for url, fetched in zip(fetch_urls, results):
if fetched.get("error"):
continue
_content = fetched.get("raw_content", "") or fetched.get("content", "")
if _content:
extract_cache_put(url, _content, fetched.get("title", ""), format=format, provider=provider.name)
return results
async def _extract_safe_urls(provider, safe_urls: List[str], format: Optional[str]) -> List[dict]:
"""Serve cache hits, fetch the rest, and merge back in ``safe_urls`` order.
The disk cache (tools/web_result_cache.py) sits AFTER the secret-URL gate, SSRF gate, and provider
resolution, and is gated per-URL on the website policy — a hit skips only the vendor call, never a
control. Policy-blocked URLs are cache misses so dispatch handles them exactly as without a cache.
Keys include provider and format, so switching either within the TTL never serves the other's content.
"""
from tools.web_result_cache import extract_cache_get
from tools.website_policy import check_website_access as _check_site
cached_results: Dict[int, Dict[str, Any]] = {}
fetch_urls: List[str] = []
fetch_positions: List[int] = []
for position, url in enumerate(safe_urls):
try:
_policy_block = _check_site(url)
except Exception: # noqa: BLE001 — policy errors fail open like dispatch
_policy_block = None
hit = extract_cache_get(url, format=format, provider=provider.name) if _policy_block is None else None
if hit is not None:
cached_results[position] = hit
else:
fetch_urls.append(url)
fetch_positions.append(position)
if not fetch_urls:
return [cached_results[i] for i in range(len(safe_urls))]
logger.info("Web extract via %s: %d URL(s)", provider.name, len(fetch_urls))
results = await _dispatch_extract(provider, fetch_urls, format)
if not cached_results:
return results
return _merge_in_order(len(safe_urls), cached_results, fetch_positions, fetch_urls, results)