Skip to contents

Retrieval-grounded corpus chat

Source: R/chat.R
sm_chat.Rd

Ask a question of a corpus using retrieval-augmented generation (RAG). The function retrieves the most relevant works from the corpus using embeddings, TF-IDF, or a hybrid method, then sends them as context to an LLM along with the user's question. The response includes inline citations keyed to work IDs. No hallucinated references are possible because citations are constrained to retrieved works.

Usage

sm_chat(
  corpus,
  question,
  provider = NULL,
  retrieve_n = 30L,
  retrieve_method = c("embedding", "tfidf", "hybrid"),
  cite = TRUE,
  max_tokens = 2000L,
  temperature = 0.2,
  call = rlang::caller_env()
)

# S3 method for class 'sm_chat_response'
print(x, ...)

Arguments

corpus

An sm_corpus object.

question

Character. The user's question in natural language.

provider

An ellmer chat provider object (e.g., from ellmer::chat_openai()). If NULL, errors informatively.

retrieve_n

Integer. Number of works to retrieve as context. Default 30.

retrieve_method

Character. Retrieval method:

  • "embedding": cosine similarity on corpus$embeddings.

  • "tfidf": TF-IDF on titles and abstracts.

  • "hybrid": weighted average of both methods.

cite

Logical. If TRUE, the prompt instructs the LLM to include inline citations using work IDs.

max_tokens

Integer. Maximum tokens for the LLM response.

temperature

Numeric. LLM temperature (0 = deterministic).

call

Caller environment for error reporting.

x

An sm_chat_response object.

...

Ignored.

Value

An sm_chat_response S3 object with fields:

answer

Character. The LLM's response.

citations

Tibble of cited work IDs and their details.

retrieved_works

Tibble of the works sent as context.

prompt_hash

SHA-256 hash of the full prompt for reproducibility.

model

Character. The model used.

timestamp

POSIXct. When the response was generated.

question

The original question.

retrieve_method

The retrieval method used.

See also

Other chat: sm_chat_render()

Examples

if (FALSE) { # \dontrun{
# Requires ellmer package and LLM API access:
# response <- sm_chat(corpus, "What methods dominate?",
#                     provider = ellmer::chat_openai())
} # }