Deprecated: answer a question using v1 mode names

Maps the old mode strings onto recipes. Unlike v1, each mode runs as an isolated pipeline, so selecting several modes no longer changes any of their answers, and mode no longer defaults to running all five.

Usage

answer_question(
  file_path,
  question,
  mode = "Chunked",
  use_parallel = FALSE,
  refine = FALSE,
  return_json = FALSE,
  client = NULL,
  ...
)

Arguments

file_path

Path to the document.

question

The question.

mode

One or more of "Retrieval", "Chunked", "Semantic", "Hierarchical", "MultiPass". Defaults to "Chunked" only.

use_parallel

Run per-chunk calls in parallel.

refine

Request chunk-level citations in the answer. v1's refine verification pass is not reproduced: it could never run, because it called a search_text() function that was never defined.

return_json

Return the answers, the comparison summary and the trace as JSON instead of the answer string.

client

A gr_client.

...

Passed to answer_document().

Value

A string, or a named list of strings when several modes are given.

See also

answer_document(), gr_compare(), gr_recipes()

Other v1 compatibility: gpt_read_chunked(), gpt_read_hierarchical(), gpt_read_multipass(), gpt_read_retrieval(), parse_text()

Examples

cl <- gr_mock_client(function(m, p) "45.2 million dollars")

# v1 style, still works, warns once.
suppressWarnings(
  answer_question(readgpt_example(), "What was revenue?", mode = "Chunked", client = cl))
#> Extracting 'annual_report.md' with the 'md' extractor.
#> Ingested 17 block(s), ~573 tokens (11 chars removed by cleaning).
#> Segmenting with 'paragraph' (cap 1200 tokens, overlap 0).
#> Reading with 'map_reduce' (all|N+logN|tree) over 1 chunk(s).
#> [1] "45.2 million dollars"

# The modern equivalent.
answer_document(readgpt_example(), "What was revenue?", "thorough",
                client = cl, return = "text")
#> Extracting 'annual_report.md' with the 'md' extractor.
#> Ingested 17 block(s), ~573 tokens (11 chars removed by cleaning).
#> Segmenting with 'paragraph' (cap 1200 tokens, overlap 120).
#> Reading with 'map_reduce' (all|N+logN|tree) over 1 chunk(s).
#> [1] "45.2 million dollars"