Perplexity web search in Hermes

Hermes Agent can use Perplexity as the backend for its web_search and web_extract tools. web_search returns ranked results from the Search API, while web_extract returns passages relevant to each URL.

This integration changes Hermes's web tools, not the model that runs the agent. You can keep your existing model provider.

The setup and verification steps below cover API-key-based standard web search and page extraction. You do not need a Perplexity API key or this setup to use the default Fast Search integration.

• A Hermes model provider configured for reasoning and writing
• A Perplexity API key for the other Perplexity search options and the extraction setup below, not for free Fast Search
• Hermes v0.21.1 (tag v2026.9.7) or later

The wizard is the recommended path. To configure the integration manually, add the key to your existing ~/.hermes/.env:

PERPLEXITY_API_KEY=pplx-your-key-here

Merge these values into the existing web section of ~/.hermes/config.yaml:

web:
  backend: perplexity
  search_backend: perplexity
  extract_backend: perplexity

search_backend and extract_backend take precedence over backend. Setting all three to perplexity prevents an older per-tool override from routing one of the calls elsewhere. If an installation already has a saved web provider, adding the API key by itself does not change that selection.

During verification, temporarily disable caching and keyless fallback in the same web section. This makes a failed Perplexity request visible instead of allowing a cached response or another provider to satisfy it.

web:
  backend: perplexity
  search_backend: perplexity
  extract_backend: perplexity
  keyless_fallback: false
  keyless_rescue: false
  cache_enabled: false

Start a fresh session with the web toolset:

hermes chat --toolsets web

Paste this smoke-test prompt:

Call web_search with query "what is a bloom filter" and limit 3.
Show the title, URL, and description for each returned result.
Do not answer from memory. If the tool fails, show the error and stop.

A successful test has both of these properties:

• Hermes shows a web_search call with the requested query.
• The tool returns usable titles, URLs, and descriptions.

Do not count a prose answer without a tool call as a successful integration test. Search results can change, so do not require specific rankings or URLs.

This tutorial uses both tools to answer a bounded engineering question: does Python 3.13 disable the GIL by default, and what should you verify before trying a free-threaded build?

Keep the verification configuration above and remain in the same Hermes session.

Paste:

I am evaluating Python 3.13 free threading for a small CPU-bound service.

Make one web_search call with query "Python 3.13 free threading"
and limit 3.

List the returned titles, URLs, and descriptions. Identify which are
official Python documentation. Do not infer that a result about another
Python version describes Python 3.13.

Stop after listing the sources. If search fails, report the error.

Inspect the returned URLs before continuing. The prompt asks Hermes to identify official sources; it does not apply an API-enforced domain filter.

Paste:

Call web_extract on exactly this URL:
https://docs.python.org/3.13/howto/free-threading-python.html

If extraction fails or returns no passages, report the error and stop.

Using only the returned content, answer:
1. Does the excerpt establish that a standard Python 3.13 build runs with
   the GIL disabled by default?
2. How can I tell whether a build supports free threading?
3. Does the excerpt give the exact function name for checking whether the
   GIL is enabled in the running process?
4. What can happen when I import an extension that does not support
   free threading?

For each answer, first quote the exact supporting passage. Do not add a
function, command, flag, or identifier unless it appears verbatim in that
passage. If a passage is missing the requested detail, say "Not established
by the retrieved excerpt." In particular, if the passage says only "the new
function" without naming it, do not supply a function name.

Perplexity extraction returns query-relevant passages, not a guaranteed complete copy of the page. It can preserve a reference such as "the new function" while omitting the function's exact name. A missing-evidence answer is valid and prevents the agent from guessing details that were omitted from the extracted text.

Paste:

Using only the passages returned by web_extract for the pinned Python
3.13 URL, write a decision brief under 250 words with these sections:

- What Python 3.13 supports
- What I need to verify in my interpreter and dependencies
- What the retrieved evidence does not establish

Cite the source URL beside each factual claim.
Do not claim that my service will run faster without a benchmark.
Do not make more tool calls, install software, or change files.

Count the walkthrough as successful when search returns usable results, extraction returns nonempty passages without a per-URL error, every factual claim is supported by those passages, and missing evidence remains explicit.

Hermes's Perplexity provider sends requests directly over HTTP. You do not need the Perplexity SDK, CLI, or an MCP server for this setup.

After verification, you can restore cache_enabled: true to reduce repeated requests. Decide separately whether to enable keyless rescue: it improves availability, but a failed Perplexity call may then be served by another provider.