Files

285 lines
11 KiB
Python

"""Perplexity web search + page snippets — plugin form.
Subclasses :class:`agent.web_search_provider.WebSearchProvider`. Two
capabilities advertised:
- ``supports_search()`` -> True (Perplexity Search API ``POST /search``)
- ``supports_extract()`` -> True (``POST /sdk/content/snippets`` — the
query-relevant page-excerpt route behind ``pplx content snippets``)
Both are sync — the underlying call is ``httpx.post(...)``.
Config keys this provider responds to::
web:
search_backend: "perplexity" # explicit per-capability
extract_backend: "perplexity" # explicit per-capability
backend: "perplexity" # shared fallback for both
Env vars::
PERPLEXITY_API_KEY=... # required for direct search and extract
PERPLEXITY_BASE_URL=... # optional override of https://api.perplexity.ai
No anonymous tier. The Nous Subscription selection serves search through
``perplexity-gateway.<TOOL_GATEWAY_DOMAIN>`` using the Nous token; a direct
key takes precedence. Managed extract stays on Firecrawl.
Extract caveat: Perplexity's only supported page-content route returns the
passages of a page relevant to a *query* (elisions marked ``…``), not the
whole page. ``web_extract`` has no query, so the URL's own path words are
used as the relevance query, which approximates "what is this page about".
Use Firecrawl / Exa / Parallel when a verbatim full-page dump is required.
"""
from __future__ import annotations
import logging
from typing import Any, Dict, List
from urllib.parse import urlparse
import httpx
from agent.web_search_provider import WebSearchProvider
from hermes_cli.version_info import get_version_info
logger = logging.getLogger(__name__)
_DEFAULT_BASE_URL = "https://api.perplexity.ai"
_KEY_URL = "https://www.perplexity.ai/account/api"
# Identify Hermes to Perplexity: the same static harness identity Hermes sends Kimi and
# OpenCode, plus Perplexity's integration header. No per-user identifier and no separate
# request; the call already carries the user's own API key.
_HEADERS = {
"HTTP-Referer": "https://hermes-agent.nousresearch.com",
"X-Title": "Hermes Agent",
"User-Agent": f"HermesAgent/{get_version_info().base_version}",
"X-Pplx-Integration": "hermes-agent",
}
# Search API hard cap for search_type=web.
_MAX_SEARCH_RESULTS = 20
# Snippet budgets (backend limits: max_tokens 1-16384, per page 1-4096).
_MAX_TOKENS = 16384
_MAX_TOKENS_PER_PAGE = 4096
def _missing_key_error() -> str:
return f"PERPLEXITY_API_KEY is not set. Get a key at {_KEY_URL}"
def _managed_gateway(token_reader=None):
"""Nous Tool Gateway config when web_search is on the managed route, else None."""
from tools import managed_tool_gateway as gw
from tools.web_tools import _managed_web_search
if not _managed_web_search():
return None
return gw.resolve_managed_tool_gateway("perplexity", token_reader=token_reader)
def _perplexity_request(endpoint: str, payload: Dict[str, Any], gateway=None) -> Dict[str, Any]:
"""POST to Perplexity or the supplied gateway; return parsed JSON.
Raises ``ValueError`` when the key is missing or on any non-2xx status,
carrying the response body so Perplexity's own error text (invalid key,
BAD_REQUEST, rate limit) reaches the model verbatim.
"""
from agent.web_search_provider import get_provider_env
api_key = get_provider_env("PERPLEXITY_API_KEY")
headers = _HEADERS
if gateway is not None:
# Nous-owned key behind the gateway: identify the harness only, not a per-user integration.
base_url, api_key, headers = gateway.gateway_origin.rstrip("/"), gateway.nous_user_token, {"User-Agent": _HEADERS["User-Agent"]}
elif api_key:
base_url = (get_provider_env("PERPLEXITY_BASE_URL") or _DEFAULT_BASE_URL).rstrip("/")
else:
raise ValueError(_missing_key_error())
url = f"{base_url}/{endpoint.lstrip('/')}"
logger.info("Perplexity %s request to %s", endpoint, url)
response = httpx.post(
url,
json=payload,
timeout=60,
headers={
"Authorization": f"Bearer {api_key}",
"Content-Type": "application/json",
**headers,
},
)
if response.status_code >= 400:
body = (response.text or "").strip()
raise ValueError(body or f"HTTP {response.status_code}")
return response.json()
def _normalize_search_results(response: Dict[str, Any]) -> Dict[str, Any]:
"""Map Search API ``{results: [{title,url,snippet,...}]}`` to the tool shape."""
web_results = []
for i, result in enumerate(response.get("results") or []):
web_results.append(
{
"title": result.get("title", "") or "",
"url": result.get("url", "") or "",
"description": result.get("snippet", "") or "",
"position": i + 1,
}
)
return {"success": True, "data": {"web": web_results}}
def _normalize_snippets(response: Dict[str, Any], urls: List[str]) -> List[Dict[str, Any]]:
"""Map ``{results: [{url,text?,tokens_count?,error?}]}`` to extract documents.
One document per requested URL, in request order. A URL the backend
omitted or flagged with ``error`` becomes a document carrying ``error``
rather than raising — a 200 does not mean every page succeeded.
"""
by_url = {r.get("url", ""): r for r in (response.get("results") or []) if isinstance(r, dict)}
documents: List[Dict[str, Any]] = []
for url in urls:
result = by_url.get(url, {})
text = result.get("text") or ""
doc: Dict[str, Any] = {
"url": url,
"title": "",
"content": text,
"raw_content": text,
"metadata": {"sourceURL": url},
}
error = result.get("error")
if error or not text:
doc["error"] = str(error) if error else "no content returned"
documents.append(doc)
return documents
def _query_for_urls(urls: List[str]) -> str:
"""Derive a relevance query from URL path words (``/bloom-filter`` -> ``bloom filter``)."""
words: List[str] = []
for url in urls:
parsed = urlparse(url)
for token in parsed.path.replace("-", " ").replace("_", " ").replace("/", " ").split():
if token.lower() not in words and not token.isdigit():
words.append(token.lower())
if not parsed.path.strip("/"):
words.append(parsed.netloc)
return " ".join(words)[:500] or " ".join(urls)[:500]
class PerplexityWebSearchProvider(WebSearchProvider):
"""Direct or managed search; direct-key content snippets for extract."""
@property
def name(self) -> str:
return "perplexity"
@property
def display_name(self) -> str:
return "Perplexity"
def is_available(self) -> bool:
"""True with a ``PERPLEXITY_API_KEY``, or on the managed route with a likely-usable Nous token."""
from agent.web_search_provider import get_provider_env
from tools.managed_tool_gateway import peek_nous_access_token
return bool(get_provider_env("PERPLEXITY_API_KEY")) or _managed_gateway(token_reader=peek_nous_access_token) is not None
def supports_search(self) -> bool:
return True
def supports_extract(self) -> bool:
return True
def search(self, query: str, limit: int = 5) -> Dict[str, Any]:
"""Execute a Perplexity Search API query.
``search_context_size: low`` keeps ``snippet`` at description length;
the default (``high``) returns multi-KB page excerpts per hit, which
belongs in ``web_extract`` rather than a results list.
"""
try:
from tools.interrupt import is_interrupted
if is_interrupted():
return {"success": False, "error": "Interrupted"}
from agent.web_search_provider import get_provider_env
from tools.web_tools import _managed_web_search
direct = bool(get_provider_env("PERPLEXITY_API_KEY"))
gateway = None if direct else _managed_gateway()
if gateway is None and not direct and _managed_web_search():
from tools.tool_backend_helpers import NOUS_MANAGED_PROVIDER, selection_error
raise ValueError(selection_error(
"web", NOUS_MANAGED_PROVIDER, "the Nous Tool Gateway is not available (not entitled or unreachable)"))
logger.info("Perplexity search: '%s' (limit=%d%s)", query, limit, ", managed" if gateway else "")
payload = {
"query": query,
"max_results": max(1, min(limit, _MAX_SEARCH_RESULTS)),
"search_context_size": "low",
}
if gateway is not None:
payload["search_type"] = "fast"
raw = _perplexity_request("search", payload, gateway)
return _normalize_search_results(raw)
except ValueError as exc:
return {"success": False, "error": str(exc)}
except Exception as exc: # noqa: BLE001 — including httpx errors
logger.warning("Perplexity search error: %s", exc)
return {"success": False, "error": f"Perplexity search failed: {exc}"}
def extract(self, urls: List[str], **kwargs: Any) -> List[Dict[str, Any]]:
"""Return query-relevant snippets for one or more URLs.
Sync — the underlying call is httpx.post(...). Per-URL failures
become items with ``error``; a missing key errors every URL.
"""
try:
from tools.interrupt import is_interrupted
if is_interrupted():
return [{"url": u, "error": "Interrupted", "title": ""} for u in urls]
logger.info("Perplexity snippets: %d URL(s)", len(urls))
raw = _perplexity_request(
"sdk/content/snippets",
{
"query": _query_for_urls(urls),
"urls": list(urls),
"max_tokens": _MAX_TOKENS,
"max_tokens_per_page": _MAX_TOKENS_PER_PAGE,
},
)
return _normalize_snippets(raw, list(urls))
except ValueError as exc:
return [{"url": u, "title": "", "content": "", "error": str(exc)} for u in urls]
except Exception as exc: # noqa: BLE001
logger.warning("Perplexity extract error: %s", exc)
return [
{"url": u, "title": "", "content": "", "error": f"Perplexity extract failed: {exc}"}
for u in urls
]
def get_setup_schema(self) -> Dict[str, Any]:
return {
"name": "Perplexity",
"badge": "paid",
"tag": (
"Perplexity Search API — ranked, date-stamped web results plus "
"query-relevant page snippets for extract."
),
"env_vars": [
{
"key": "PERPLEXITY_API_KEY",
"prompt": "Perplexity API key",
"url": _KEY_URL,
},
],
"web_tier": "paid",
}