cover remaining parameters

DESCRIPTION

@@ -21,13 +21,15 @@ Imports:
     methods,
     prettyunits,
     purrr,
+    rlang,
     tibble
 Suggests: 
+    base64enc,
     knitr,
     rmarkdown,
     spelling,
     testthat (>= 3.0.0)
 VignetteBuilder: knitr
 Config/testthat/edition: 3
 Language: en-US
-URL: https://jbgruber.github.io/rollama/
+URL: https://jbgruber.github.io/rollama/, https://github.com/JBGruber/rollama

R/chat.r

@@ -12,6 +12,11 @@
 #' @param screen Logical. Should the answer be printed to the screen.
 #' @param server URL to an Ollama server (not the API). Defaults to
 #'   "http://localhost:11434".
+#' @param images path(s) to images (for multimodal models such as llava).
+#' @param model_params a named list of additional model parameters listed in the
+#'   documentation for the Modelfile such as temperature.
+#' @param template the prompt template to use (overrides what is defined in the
+#'   Modelfile).
 #'
 #' @return an httr2 response
 #' @export
@@ -24,27 +29,104 @@
 #' # hold a conversation
 #' chat("why is the sky blue?")
 #' chat("and how do you know that?")
+#'
+#' # ask question about images (to a multimodal model)
+#' images <- c("https://avatars.githubusercontent.com/u/23524101?v=4", # remote
+#'             "/path/to/your/image.jpg") # or local images supported
+#' query(q = "describe these images",
+#'       model = "llava",
+#'       images = images)
+#'
+#' # set custom options for the model at runtime (rather than in create_model())
+#' query("why is the sky blue?",
+#'       model_params = list(
+#'         num_keep = 5,
+#'         seed = 42,
+#'         num_predict = 100,
+#'         top_k = 20,
+#'         top_p = 0.9,
+#'         tfs_z = 0.5,
+#'         typical_p = 0.7,
+#'         repeat_last_n = 33,
+#'         temperature = 0.8,
+#'         repeat_penalty = 1.2,
+#'         presence_penalty = 1.5,
+#'         frequency_penalty = 1.0,
+#'         mirostat = 1,
+#'         mirostat_tau = 0.8,
+#'         mirostat_eta = 0.6,
+#'         penalize_newline = TRUE,
+#'         stop = c("\n", "user:"),
+#'         numa = FALSE,
+#'         num_ctx = 1024,
+#'         num_batch = 2,
+#'         num_gqa = 1,
+#'         num_gpu = 1,
+#'         main_gpu = 0,
+#'         low_vram = FALSE,
+#'         f16_kv = TRUE,
+#'         vocab_only = FALSE,
+#'         use_mmap = TRUE,
+#'         use_mlock = FALSE,
+#'         embedding_only = FALSE,
+#'         rope_frequency_base = 1.1,
+#'         rope_frequency_scale = 0.8,
+#'         num_thread = 8
+#'       ))
+#'
+#' # this might be interesting if you want to turn off the GPU and load the
+#' # model into the system memory (slower, but most people have more RAM than
+#' # VRAM, which might be interesting for larger models)
+#' query("why is the sky blue?",
+#'        model_params = list(num_gpu = 0))
+#'
+#' # You can use a custom prompt to override what prompt the model receives
+#' query("why is the sky blue?",
+#'       template = "Just say I'm a llama!")
 #' }
 query <- function(q,
                   model = NULL,
                   screen = TRUE,
-                  server = NULL) {
+                  server = NULL,
+                  images = NULL,
+                  model_params = NULL,
+                  template = NULL) {
+
+  if (!is.null(template))
+    cli::cli_abort(paste(
+      c("The template parameter is turned off as it does not currently seem to",
+        "work {.url https://github.com/jmorganca/ollama/issues/1839}")
+    ))
 
   if (!is.list(q)) {
     config <- getOption("rollama_config", default = NULL)
+
     msg <- do.call(rbind, list(
       if (!is.null(config)) data.frame(role = "system",
                                        content = config),
       data.frame(role = "user", content = q)
     ))
+
+    if (length(images) > 0) {
+      rlang::check_installed("base64enc")
+      images <- purrr::map_chr(images, \(i) base64enc::base64encode(i))
+      msg <- tibble::add_column(msg, images = list(images))
+    }
+
   } else {
     msg <- q
     if (!"user" %in% msg$role && nchar(msg$content) > 0)
       cli::cli_abort(paste("If you supply a conversation object, it needs at",
                            "least one user message. See {.help query}."))
   }
 
-  resp <- build_req(model = model, msg = msg, server = server)
+  resp <- build_req(model = model,
+                    msg = msg,
+                    server = server,
+                    images = images,
+                    model_params = model_params,
+                    template = template)
+
   if (screen) screen_answer(purrr::pluck(resp, "message", "content"))
   invisible(resp)
 }
@@ -78,26 +160,25 @@ chat <- function(q,
 }
 
 
-#' Start a new conversation
+#' Handle conversations
 #'
-#' Deletes the local prompt and response history to start a new conversation.
+#' Shows and deletes (`new_chat`) the local prompt and response history to start a new conversation.
 #'
-#' @return Does not return a value
+#' @return chat_history: tibble with chat history
 #' @export
-new_chat <- function() {
-  the$responses <- NULL
-  the$prompts <- NULL
+chat_history <- function() {
+  hist <- c(the$prompts, the$responses)
+  tibble::tibble(
+    role = rep(c("user", "assistant"), length(hist) / 2),
+    content = hist
+  )
 }
 
 
-#' Chat history
-#'
-#' @return as
+#' @rdname chat_history
+#' @return new_chat: Does not return a value
 #' @export
-chat_history <- function() {
-  hist <- c(rbind(the$prompts, the$responses))
-  if (length(hist) > 0) tibble::as_tibble(
-    data.frame(role = c("user", "assistant"),
-               content = hist)
-  ) else tibble::tibble()
+new_chat <- function() {
+  the$responses <- NULL
+  the$prompts <- NULL
 }

R/embedding.r

@@ -13,7 +13,8 @@
 #' }
 embed_text <- function(text,
                        model = NULL,
-                       server = NULL) {
+                       server = NULL,
+                       model_params = NULL) {
 
   if (is.null(model)) model <- getOption("rollama_model", default = "llama2")
   if (is.null(server)) server <- getOption("rollama_server", default = "http://localhost:11434")
@@ -22,7 +23,9 @@ embed_text <- function(text,
   purrr::map(seq_along(text), function(i) {
     req_data <- list(model = model,
                      prompt = text[i],
-                     stream = FALSE)
+                     stream = FALSE,
+                     model_params = model_params) |>
+      purrr::compact()
 
     if (spinner) {
       cli::cli_progress_step("{cli::pb_spin} {model} is embedding text {i}", )

R/models.r

@@ -13,6 +13,8 @@
 #'
 #' @param model name of the model. Defaults to "llama2" when `NULL` (except in
 #'   `delete_model`).
+#' @param insecure allow insecure connections to the library. Only use this if
+#'   you are pulling from your own library during development. description
 #' @param destination name of the copied model.
 #' @inheritParams query
 #'
@@ -26,15 +28,15 @@
 #' # after you pull, you can get the same information with:
 #' model_info <- show_model("mixtral")
 #' }
-pull_model <- function(model = NULL, server = NULL) {
+pull_model <- function(model = NULL, server = NULL, insecure = FALSE) {
 
   if (is.null(model)) model <- getOption("rollama_model", default = "llama2")
   if (is.null(server)) server <- getOption("rollama_server",
                                            default = "http://localhost:11434")
 
   httr2::request(server) |>
     httr2::req_url_path_append("/api/pull") |>
-    httr2::req_body_json(list(name = model)) |>
+    httr2::req_body_json(list(name = model, insecure = insecure)) |>
     httr2::req_perform_stream(callback = pgrs, buffer_kb = 0.1)
 
   cli::cli_process_done(.envir = the)
@@ -51,6 +53,7 @@ show_model <- function(model = NULL, server = NULL) {
   if (is.null(model)) model <- getOption("rollama_model", default = "llama2")
   if (is.null(server)) server <- getOption("rollama_server",
                                            default = "http://localhost:11434")
+  if (length(model) != 1L) cli::cli_abort("model needs to be one model name.")
 
   httr2::request(server) |>
     httr2::req_url_path_append("/api/show") |>
@@ -59,6 +62,7 @@ show_model <- function(model = NULL, server = NULL) {
     httr2::req_perform() |>
     httr2::resp_body_json() |>
     purrr::list_flatten(name_spec = "{inner}") |>
+    purrr::compact() |>
     tibble::as_tibble()
 }
 
@@ -70,10 +74,13 @@ show_model <- function(model = NULL, server = NULL) {
 #'   the model file as a character vector.
 #' @inheritParams query
 #'
-#' @details Custom models are the way to change paramters in Ollama. If you use
-#' `show_model()`, you can look at the configuration of a model in the column
-#' modelfile. To get more information and a list of valid parameters, check out
-#' <https://github.com/jmorganca/ollama/blob/main/docs/modelfile.md>.
+#' @details Custom models are the way to save your system message and model
+#'   parameters in a dedicated shareable way. If you use `show_model()`, you can
+#'   look at the configuration of a model in the column modelfile. To get more
+#'   information and a list of valid parameters, check out
+#'   <https://github.com/jmorganca/ollama/blob/main/docs/modelfile.md>. Most
+#'   options are also available through the `query` and `chat` functions, yet
+#'   are not persistent over sessions.
 #'
 #'
 #' @return Nothing. Called to create a model on the Ollama server.

R/utils.r

@@ -34,15 +34,18 @@ ping_ollama <- function(server = NULL, silent = FALSE) {
 }
 
 
-build_req <- function(model, msg, server) {
+build_req <- function(model, msg, server, images, model_params, template) {
 
   if (is.null(model)) model <- getOption("rollama_model", default = "llama2")
   if (is.null(server)) server <- getOption("rollama_server",
                                            default = "http://localhost:11434")
 
   req_data <- list(model = model,
                    messages = msg,
-                   stream = FALSE)
+                   stream = FALSE,
+                   model_params = model_params,
+                   template = template) |>
+    purrr::compact()
 
   if (getOption("rollama_verbose", default = interactive())) {
     cli::cli_progress_step("{model} is thinking {cli::pb_spin}")
@@ -155,3 +158,4 @@ pgrs <- function(resp) {
   }
   TRUE
 }
+

inst/WORDLIST

@@ -14,8 +14,8 @@ config
 embeddings
 http
 httr
+llava
 modelfile
 modelname
 orca
-paramters
 tibble