Package index • readgpt

Ask a question

Start here. One call reads a document and answers a question, and the answer says what it rests on and whether anything went wrong.

answer_document(): Answer a question about a document
gr_compare(): Run several recipes over one document and compare them
gr_recipes(): Ready-made recipes
gr_recipe(): Bind an ingestion, segmentation and reading configuration together
gr_answer: The result of one reading run
is_not_found(): Did the model report that the document does not contain the answer?
readgpt_example(): Path to the bundled example document

Connect to a model

Clients for OpenAI-compatible endpoints, any provider ellmer supports, any function you write, and an offline stand-in for practice and tests.

gr_client(): Construct a model client
gr_api_key(): Resolve the API key
gr_ellmer_client(): Read documents through an ellmer chat
gr_backend_client(): Use any function as the model transport
gr_mock_client(): A deterministic offline client for tests, demos and dry runs
gr_call(): Call a model
gr_result: The result of one model call
gr_models(): List every known model
gr_model_info(): Look up a model's capabilities
gr_model_limits(): Context and output limits for a model
gr_register_model(): Register a model (or override a built-in entry)

Get text out of files

Extraction, cleaning and a survey of a folder before anything is spent. See vignette("ingest").

gr_ingest(): Ingest a document into cleaned, provenance-bearing text blocks
gr_ingest_spec(): Describe an ingestion configuration
gr_document: An ingested document
gr_extractors(): List registered extractors
gr_cleaners(): List registered cleaners
gr_clean(): Run the cleaning pipeline over a character vector
gr_inventory(): What is in a folder, before you read any of it

Cut documents into chunks

gr_segment(): Segment a document into chunks
gr_segment_spec(): Describe a segmentation configuration
gr_segmenters(): List registered segmentation strategies
gr_chunks: A set of document chunks
gr_chunk_stats(): Summary statistics for a chunk set

Read the chunks

The reading strategies, the settings that choose and place chunks, and the check on quoted evidence. See vignette("readers").

gr_read(): Read chunks and answer a question
gr_read_spec(): Describe a reading configuration
gr_readers(): List registered reading strategies
gr_reader_signature(): The traversal signature of a reader
gr_embed(): Embed texts
gr_embedders(): List registered embedding backends
gr_verify_evidence(): Check that quoted evidence really is in the document

Many documents

gr_read_many(): Ask one question of many documents

Systematic and literature reviews

From a search export to screened studies, an extraction table, a written review and an audit report, with the screening measured against a hand-screened sample.

gr_protocol(): Write down what a review is looking for
gr_protocols(): Protocols that ship with the package, and any you have registered
gr_protocol_save() gr_protocol_read(): Save a protocol to a file, and read one back
gr_register_protocol(): Register a protocol
gr_search(): Record how the search was run
gr_records(): Read a search export, and say what the search found
gr_screen(): Decide which documents a review should read
gr_reference(): Draw a sample to screen by hand
gr_calibrate(): Measure the screener against a hand-screened sample
`[`(<gr_reference_frame>): Subsetting a reference frame gives a plain data frame.
gr_fields(): Build an extraction schema
gr_field(): Describe one field of an extraction schema
gr_extract(): Extract a typed schema from many documents
gr_claims(): The claims a table of studies supports
gr_outline(): Derive a review's sections from its claims
gr_gaps(): What a body of work does not cover
gr_synthesise(): Write a review from an extraction table
gr_flow(): Count what happened to every document
gr_audit_report(): Write the run out as an auditable report

Cost, tokens and tracing

gr_options(): Get or set package options
gr_estimate_cost(): Estimate the USD cost of a set of calls
gr_count_tokens(): Count tokens in text
gr_tokenizer(): The active tokenizer
gr_set_tokenizer(): Register or inspect the active tokenizer
gr_truncate_tokens(): Truncate text to at most n tokens
gr_budget(): Compute a usable input-token budget for one model call
gr_trace() as.data.frame(<gr_trace>): Create a run trace
gr_trace_summary(): Summarise a trace
gr_trace_cost(): What a run actually cost
gr_trace_save(): Save a trace to a file
as_json(): Serialise an object to JSON

Caching and replaying runs

gr_cache(): A response cache
gr_cache_client(): Attach a cache to a client
gr_cache_stats(): Cache statistics
gr_cache_clear(): Delete every entry in a cache
gr_replay_client(): A client that answers from a recorded run

Extending readgpt

Every axis is a registry. An addition is used exactly like a built-in.

gr_register_extractor(): Register a document extractor
gr_register_cleaner(): Register a cleaning step
gr_register_segmenter(): Register a segmentation strategy
gr_register_reader(): Register a reading strategy
gr_register_embedder(): Register an embedding backend
new_chunks(): Build a gr_chunks, the object every segmenter must return
new_answer(): Build a gr_answer, the object every reader must return

Superseded

The entry points of the first version. They still work and warn once; new code should use answer_document(), gr_ingest() and gr_segment().

answer_question(): Deprecated: answer a question using v1 mode names
parse_text(): Deprecated: parse a document into text chunks
gpt_read_chunked(): Deprecated: chunk-by-chunk reading
gpt_read_hierarchical(): Deprecated: hierarchical reading
gpt_read_multipass(): Deprecated: multi-pass reading
gpt_read_retrieval(): Deprecated: evidence-extraction reading