R functions that answer questions about text

The ten functions work inside dplyr pipelines. Read the answer from $value.

Install

Channel
`install.packages("thinkthen")`

Set THINKTHEN_API_KEY in the environment.

A first call

library(thinkthen)

question <- "Does the customer ask for a refund?"

broken <- "Please refund my order. It arrived broken."
is_refund <- tt_decide(question, broken)$value
stopifnot(identical(is_refund, TRUE))

send_back <- "I want to send this back."
is_refund <- tt_decide(
  question,
  send_back,
  threshold = "0.2:0.8"
)$value
stopifnot(identical(is_refund, NA))

Not sure comes back as NA. Branch on it and send those cases to a person. Each function page shows its R call.

To apply one question to a list or a stream, read Functional patterns.

Data frames

A verb inside mutate() answers a whole column in one call. dplyr's filter() drops NA rows, so a not-sure answer leaves the pipeline on its own.

library(dplyr)
library(thinkthen)

question <- "Does the customer ask for a refund?"
messages <- tibble(
  body = c(
    "Please refund my order. It arrived broken.",
    "I want to send this back."
  )
)

answered <- messages |>
  mutate(
    is_refund = tt_decide(
      question,
      body,
      threshold = "0.2:0.8"
    )$value
  )
stopifnot(identical(answered$is_refund, c(TRUE, NA)))

refunds <- answered |> filter(is_refund)
stopifnot(identical(refunds$body, messages$body[1]))

Run facts

$facts counts this call.

Errors

Each failure arrives as an R condition named for its kind, such as thinkthen_usage. Each condition carries retryable.

Settings and recording

tt_engine() takes each setting as an argument and reads the environment for the rest. record = writes a recording to a folder. replay = answers from that recording with no connection. The Settings page lists every setting.

Jev, Liquid d1 and Ollama from R

R reads its backend from the environment. Set THINKTHEN_BACKEND to the name below. Set the backend's key variable in the shell, and the same sample asks that backend.

Backend	Name	Key variable
TypeSafe Jev	`typesafe`	`TYPESAFE_API_KEY`
Liquid d1	`liquid`	`LIQUIDAI_API_KEY`, then `LIQUID_API_KEY`
Ollama	`ollama`	`OLLAMA_API_KEY`. A local address needs no key.

Ask TypeSafe Jev.

export THINKTHEN_BACKEND=typesafe
export TYPESAFE_API_KEY=...

Ask Liquid d1.

export THINKTHEN_BACKEND=liquid
export LIQUIDAI_API_KEY=...

Ask Ollama. This block names a second port, 11535, as the site's runs do. On the default port, leave THINKTHEN_BASE_URL out.

export THINKTHEN_BACKEND=ollama
export THINKTHEN_BASE_URL=http://localhost:11535/v1

library(thinkthen)

question <- "Does the customer ask for a refund?"
texts <- c(
  "Please refund my order. It arrived broken.",
  "Thanks for the quick help yesterday!"
)
is_refund <- tt_decide(question, texts)$value
stopifnot(identical(is_refund, c(TRUE, FALSE)))

The configuration file's backend names a default backend for every run. The Configuration page shows the file.

OpenAI Decisions API from R

OpenAI announced its Decisions API on 2026-09-29 in its DevDay 2026 recap. OpenAI has published no address, schema or price, so ThinkThen cannot call it. The OpenAI Decisions API page says what ThinkThen knows.

What is different in R

A column goes in and the answered column is in $value.

On GitHub: github.com/botassembly/thinkthen