Perplexed.Pipeline (Perplexed v0.2.0)

Copy Markdown View Source

One-shot web answering, in fixed steps:

  1. Plan - one model call turns the question into a few search queries (skipped when the caller passes :queries).
  2. Search - every query runs concurrently; hits are de-duplicated by URL.
  3. Rank - one model call scores each hit's relevance from 0 to 1, seeing only its title, URL and description. Hits at or above :min_relevance are kept, best first.
  4. Scrape - the best :max_sources kept hits are scraped concurrently; an unreadable page is replaced by the next kept hit, and a page another source already covers is skipped.
  5. Summarize - one model call writes the answer from the numbered sources, as text with [n] citations or as an object matching :schema.

The result is a Perplexed.Answer carrying the answer, the sources it was written from, and every hit considered with why it was used or not.

Untrusted content

Search results and page text are written by third parties. They reach the model inside <result> and <source> tags, with a standing instruction to treat that text as data, never as instructions. That lowers, but does not remove, the risk of a page steering its own rank or the answer; treat answers, and especially structured answers you act on, accordingly.

Summary

Types

t()

One run's resolved options and accumulated state.

Functions

run(query, opts \\ [])

Answers query from the web.

Types

t()

@type t() :: %Perplexed.Pipeline{
  answer: term(),
  candidates: term(),
  cited: term(),
  explicit_queries: term(),
  generate_object: term(),
  generate_text: term(),
  hits_per_query: term(),
  instructions: term(),
  max_queries: term(),
  max_sources: term(),
  min_relevance: term(),
  model: term(),
  model_options: term(),
  object_options: term(),
  page_chars: term(),
  provider: term(),
  queries: term(),
  query: term(),
  schema: term(),
  scrape_options: term(),
  search_options: term(),
  sources: term(),
  task_supervisor: term(),
  trace: term(),
  use_search_content: term(),
  validation_retries: term()
}

One run's resolved options and accumulated state.

Functions

run(query, opts \\ [])

@spec run(String.t(), keyword()) ::
  {:ok, Perplexed.Answer.t()} | {:error, Perplexed.error()}

Answers query from the web.

Options

  • :model - a ReqLLM model spec, default config :perplexed, model:. Any provider ReqLLM supports works.
  • :model_options - ReqLLM options for every model call, overriding the built-in max_tokens: 4000 and receive_timeout:; :object_options - extra options for structured-output calls.
  • :schema - a Zoi schema; the answer becomes a validated object.
  • :instructions - system instructions for the summary call, replacing the built-in ones (the untrusted-content note is always kept).
  • :queries - search queries to run instead of planning them.
  • :max_queries - planned queries, default 3.
  • :provider, :search_options - the search provider spec and its options; :hits_per_query defaults to 8.
  • :max_sources - pages to read and summarize, default 5.
  • :min_relevance - lowest ranking score kept, default 0.5.
  • :scrape_options - options for Perplexed.Scrape.batch/2, such as :ladder.
  • :page_chars - characters of each page given to the summary, default 12000.
  • :use_search_content - read the page text a search provider returned with a hit (DeepSeek does) instead of scraping it, when that text passes Perplexed.Verdict; default true. Such a source's tier is :search.
  • :validation_retries - schema-mismatch retries, default 1.
  • :task_supervisor - default Perplexed.TaskSupervisor.
  • :generate_object, :generate_text - stand-ins for ReqLLM.generate_object/4 and ReqLLM.generate_text/3, for tests.