Answer a question about a document

One recipe, one pipeline: ingest, segment, read. The answer and its full trace come from a single run, so the trace always explains the answer you got.

Usage

answer_document(
  source,
  question,
  recipe = "auto",
  client = NULL,
  return = c("answer", "text", "json"),
  trace = NULL,
  ...
)

Arguments

source

File path, web address, or raw text; see gr_ingest().

question

The question.

recipe

A gr_recipe, a recipe name from gr_recipes(), a reader name, or a named list of ingest/segment/read. The default, "auto", picks "fast" or "thorough" from the document's length; see "Choosing the recipe" below.

client

A gr_client; one is built from options when omitted.

return

"answer" (a gr_answer), "text" (the string), or "json" (answer plus trace, serialised).

trace

Optional gr_trace to accumulate into.

...

Convenience overrides applied to the recipe: any gr_read_spec, gr_segment_spec or gr_ingest_spec field (for example model, max_tokens, top_k, clean). Unknown names raise an error instead of being silently discarded.

Value

Depends on return. The gr_answer carries $partial; check it before trusting $answer.

Choosing the recipe

With recipe = "auto" the document is ingested first, and its length decides how it is read. "fast" sends the whole document in one request. It is used for a document of at most 50,000 tokens that also fills no more than half of the room one request leaves for the document, which is less on a model with a small context window. Anything longer is read with "thorough": one request per chunk, then the requests that combine their answers. Both send every chunk, so the choice changes the number of requests and the cost, not how much of the document is read.

The room is measured for the recipe's model and, unless model is passed in ..., for the client's model as well, since a client built for another model may be what answers. A model whose limits readgpt has to guess (see gr_model_info()) always gets "thorough"; register its real limits with gr_register_model().

The answer's recipe names the recipe used, notes$auto_recipe records that "auto" chose it, and the trace has an auto_recipe step with the token count and the limit the choice was made on. A gr_replay_client() repeats the recorded choice rather than making it again. Overrides in ... apply to whichever recipe is chosen. gr_read_many(), gr_compare() and the review functions need one fixed recipe and refuse "auto".

See also

gr_recipes() for the built-in pipelines, gr_compare() to run several, gr_answer for the returned object, gr_options() for the cost and call caps

Examples

cl <- gr_mock_client(function(m, p) "The answer is 42.")
txt <- "Chapter one.\n\nThe answer to the great question is 42, as recorded."
answer_document(txt, "What is the answer?", "fast", client = cl, return = "text")
#> Ingested 2 block(s), ~24 tokens (0 chars removed by cleaning).
#> Segmenting with 'paragraph' (cap 4000 tokens, overlap 0).
#> Reading with 'stuff' (all|1|none) over 1 chunk(s).
#> [1] "The answer is 42."