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 a chat provider 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 = lifecycle::deprecated(),
temperature = lifecycle::deprecated(),
call = rlang::caller_env()
)
# S3 method for class 'sm_chat_response'
print(x, ...)Arguments
- corpus
An
sm_corpusobject.- question
Character. The user's question in natural language.
- provider
An ellmer chat provider object (e.g., from
ellmer::chat_openai()). IfNULL, errors informatively. Each request runs in a fresh copy: the object passed is unchanged, and its earlier turns and system prompt are not used.- retrieve_n
Integer. Number of works to retrieve as context. Default
30.- retrieve_method
Character. Retrieval method:
"embedding": cosine similarity oncorpus$embeddings."tfidf": TF-IDF on titles and abstracts."hybrid": weighted average of both methods.
- cite
Logical. If
TRUE, the prompt asks the chat provider for inline citations using work IDs.- max_tokens, temperature
Deprecated; these arguments never affected requests. Set response parameters on the chat object instead.
- call
Caller environment for error reporting.
- x
An
sm_chat_responseobject.- ...
Ignored.
Value
An sm_chat_response S3 object with fields:
- answer
Character. The chat provider'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{
# Not run: needs the 'ellmer' package and credentials for a chat provider.
corpus <- sm_example_corpus()
response <- sm_chat(corpus, "What methods dominate?",
provider = ellmer::chat_openai())
response
} # }