Merge pull request #22 from bakaburg1/Dev

Improved LLM Handling and Transcript Management

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: minutemaker
 Title: GenAI-based meeting and conferences minutes generator
-Version: 0.9.0
+Version: 0.10.0
 Authors@R: 
     person("Angelo", "D'Ambrosio", , "a.dambrosioMD@gmail.com", role = c("aut", "cre"),
            comment = c(ORCID = "0000-0002-2045-5155"))

LICENSE

@@ -1,21 +1,2 @@
-MIT License
-
-Copyright (c) 2023 Angelo D'Ambrosio
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
+YEAR: 2023
+COPYRIGHT HOLDER: Angelo D'Ambrosio

LICENSE.md

@@ -1,6 +1,6 @@
 # MIT License
 
-Copyright (c) 2023 minutemaker authors
+Copyright (c) 2023 Angelo D'Ambrosio
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

NAMESPACE

@@ -10,10 +10,10 @@ export(generate_recording_details)
 export(get_prompts)
 export(import_transcript_from_file)
 export(infer_agenda_from_transcript)
-export(interrogate_llm)
 export(merge_transcripts)
 export(parse_transcript_json)
 export(perform_speech_to_text)
+export(prompt_llm)
 export(run_in_terminal)
 export(set_prompts)
 export(speech_to_summary_workflow)

NEWS.md

@@ -1,3 +1,22 @@
+# minutemaker 0.10.0
+
+#### Enhancements
+- **Change interrogate_llm to prompt_llm**: Renamed the `interrogate_llm` function to `prompt_llm` to better reflect its purpose of generating prompts for language models (Commit: [2f9eeddd]).
+- **Model Logging for LLM Requests**: Added a `log_request` parameter to `use_openai_llm`, `use_azure_llm`, and `use_custom_llm` functions to log the specific endpoint and model being used (Commit: [7e85f2f]).
+- **Handling Long LLM Responses**: Improved the handling of LLM responses that exceed the output token limit. Users are now prompted to decide how to proceed, and incomplete answers are saved to a file for reference (Commit: [18cfada]).
+- **Model Parameter for Custom LLM APIs**: Added a `model` parameter to LLM calls to allow specifying a model for custom APIs with multiple models (Commit: [cd4227b]).
+
+#### Fixes
+- **Restore Correct Speaker Extraction for Webex VTT**: Fixed the parsing of Webex VTT files which was broken by the implementation of MS Teams VTT parsing (Commit: [d189980]).
+- **Remove Newlines from JSON Output**: Fixed an issue where some custom LLMs produced invalid JSON with newlines, causing errors during parsing (Commit: [e9e578a]).
+- **Support JSON Mode for Custom LLMs**: Ensured that most custom LLMs now support JSON mode by keeping the JSON mode option in the call (Commit: [f4df24c]).
+
+#### Documentation
+- **Improve Code Comments and Error Handling**: Enhanced code comments and error handling for better clarity and maintenance (Commit: [4b689ff]).
+
+#### Summary
+This pull request introduces several enhancements to the LLM handling, including logging the model being used, better management of long responses, and support for specifying models in custom APIs. It also includes fixes for speaker extraction in Webex VTT files and handling of JSON outputs from custom LLMs. Additionally, code comments and error handling have been improved for better clarity.
+
 # minutemaker 0.9.0
 
 ### Improve agenda review and add custom LLM support

R/LLM_calls.R

@@ -120,7 +120,7 @@ process_messages <- function(messages) {
 
 }
 
-#' Interrogate Language Model
+#' Interrogate a Language Model
 #'
 #' This function sends requests to a specified language model provider (OpenAI,
 #' Azure, or a locally running LLM server) and returns the response. It handles
@@ -149,12 +149,12 @@ process_messages <- function(messages) {
 #'
 #' @examples
 #' \dontrun{
-#' response <- interrogate_llm(
+#' response <- prompt_llm(
 #'  messages = c(user = "Hello there!"),
 #'  provider = "openai")
 #'  }
 #'
-interrogate_llm <- function(
+prompt_llm <- function(
     messages = NULL,
     provider = getOption("minutemaker_llm_provider"),
     params = list(
@@ -183,7 +183,7 @@ interrogate_llm <- function(
 
   body$messages <- messages
 
-  # Force the LLM to answer in JSON format (only openai and azure)
+  # Force the LLM to answer in JSON format (not all models support this)
   if (force_json) {
     body$response_format <- list("type" = "json_object")
   }
@@ -192,7 +192,8 @@ interrogate_llm <- function(
   llm_fun <- paste0("use_", provider, "_llm")
 
   if (!exists(llm_fun, mode = "function")) {
-    stop("Unsupported LLM provider.")
+    stop("Unsupported LLM provider.
+         You can set it project-wide using the minutemaker_llm_provider option.")
   }
 
   llm_fun <- get(llm_fun)
@@ -252,13 +253,62 @@ interrogate_llm <- function(
     ) |> message()
   }
 
-  answer <- parsed$choices[[1]]
-
-  if (answer$finish_reason == "length") {
-    stop("Answer exhausted the context window!")
-  }
-
-  answer$message$content
+  # Return the response
+  purrr::imap_chr(parsed$choices, \(ans, i) {
+    ans_content <- ans$message$content
+
+    # Manage the case when the answer is cut off due to exceeding the
+    # output token limit
+    if (ans$finish_reason == "length") {
+      i <- if (length(parsed$choices) > 1) paste0(" ", i, " ") else " "
+
+      warning("Answer", i, "exhausted the context window!")
+
+      file_name <- paste0("output_", Sys.time(), ".txt")
+
+      warning(
+        "Answer", i, "exhausted the context window!\n",
+        "The answer has been saved to a file: ", file_name
+      )
+
+      readr::write_lines(ans_content, file_name)
+
+      choice <- utils::menu(
+        c(
+          "Try to complete the answer",
+          "Keep the incomplete answer",
+          "Stop the process"),
+        title = "How do you want to proceed?"
+      )
+
+      if (choice == 1) {
+        # Ask the model to continue the answer
+        messages_new <- c(
+          messages,
+          list(list(
+            role = "assistant",
+            content = ans_content
+          )),
+          list(list(
+            role = "user",
+            content = "continue"
+          ))
+        )
+
+        ans_new <- prompt_llm(
+          messages_new, provider = provider, params = params,
+          force_json = force_json,
+          log_request = log_request, ...
+        )
+
+        return(paste0(ans_content, ans_new))
+      } else if (choice == 2) {
+        return(ans_content)
+      } else {
+        stop("The process has been stopped.")
+      }
+    } else ans_content
+  })
 }
 
 #' Use OpenAI Language Model
@@ -269,13 +319,16 @@ interrogate_llm <- function(
 #' @param body The body of the request.
 #' @param model Model identifier for the OpenAI API. Obtained from R options.
 #' @param api_key API key for the OpenAI service. Obtained from R options.
+#' @param log_request A boolean to log the request time. Can be set up globally
+#'   using the `minutemaker_log_requests` option, which defaults to TRUE.
 #'
 #' @return The function returns the response from the OpenAI API.
 #'
 use_openai_llm <- function(
     body,
     model = getOption("minutemaker_openai_model_gpt"),
-    api_key = getOption("minutemaker_openai_api_key")
+    api_key = getOption("minutemaker_openai_api_key"),
+    log_request = getOption("minutemaker_log_requests", TRUE)
 ) {
 
   if (is.null(api_key) || is.null(model)) {
@@ -285,6 +338,10 @@ use_openai_llm <- function(
          "minutemaker_open_api_key options.")
   }
 
+  if (log_request) {
+    message("Interrogating OpenAI: ", model, "...")
+  }
+
   body$model = model
 
   # Prepare the request
@@ -315,14 +372,17 @@ use_openai_llm <- function(
 #'   options.
 #' @param api_version API version for the Azure language model service. Obtained
 #'   from R options.
+#' @param log_request A boolean to log the request time. Can be set up globally
+#'   using the `minutemaker_log_requests` option, which defaults to TRUE.
 #'
 #' @return The function returns the response from the Azure API.
 use_azure_llm <- function(
     body,
     deployment_id = getOption("minutemaker_azure_deployment_gpt"),
     resource_name = getOption("minutemaker_azure_resource_gpt"),
     api_key = getOption("minutemaker_azure_api_key_gpt"),
-    api_version = getOption("minutemaker_azure_api_version")
+    api_version = getOption("minutemaker_azure_api_version"),
+    log_request = getOption("minutemaker_log_requests", TRUE)
 ) {
 
   if (is.null(resource_name) || is.null(deployment_id) ||
@@ -337,6 +397,12 @@ use_azure_llm <- function(
     )
   }
 
+  if (log_request) {
+    message(
+      "Interrogating Azure OpenAI: ", resource_name, "/", deployment_id,
+      " (", api_version, ")...")
+  }
+
   # Prepare the request
   httr::POST(
     url = paste0(
@@ -361,15 +427,21 @@ use_azure_llm <- function(
 #' @param body The body of the request.
 #' @param endpoint The local endpoint for the language model service. Can be
 #'   obtained from R options.
+#' @param model Model identifier for the custom API, if needed (some API have
+#'   one model per endpoint, some multiple ones). Obtained from R options.
 #' @param api_key Optional API key for the custom language model services that
 #'   require it. Obtained from R options.
+#' @param log_request A boolean to log the request time. Can be set up globally
+#'   using the `minutemaker_log_requests` option, which defaults to TRUE.
 #'
 #' @return The function returns the response from the local language model
 #'   endpoint.
 use_custom_llm <- function(
     body,
     endpoint = getOption("minutemaker_custom_endpoint_gpt"),
-    api_key = getOption("minutemaker_custom_api_key")
+    model = getOption("minutemaker_custom_model_gpt", NULL),
+    api_key = getOption("minutemaker_custom_api_key"),
+    log_request = getOption("minutemaker_log_requests", TRUE)
 ) {
 
   if (is.null(endpoint)) {
@@ -379,7 +451,13 @@ use_custom_llm <- function(
     )
   }
 
-  body$response_format <- NULL
+  if (log_request) {
+    message("Interrogating custom LLM: ", endpoint, "/", model, "...")
+  }
+
+  if (!is.null(model)) {
+    body$model = model
+  }
 
   # Prepare the request
   httr::POST(
@@ -393,4 +471,3 @@ use_custom_llm <- function(
   )
 
 }
-

R/data_management.R

@@ -557,8 +557,9 @@ format_summary_tree <- function(
     agenda_element$title <- ifelse(is.null(
       agenda_element$title), "", agenda_element$title)
 
-    output_piece <- stringr::str_glue_data(agenda_element,
-                                           "Session: {session};
+    output_piece <- stringr::str_glue_data(
+      agenda_element,
+      "Session: {session};
     Title: {title};
     Speakers: {speakers};
     Moderators: {moderators};
@@ -597,8 +598,8 @@ format_summary_tree <- function(
 #'
 #' @export
 format_agenda <- function(
-  agenda,
-  event_start_time = getOption("minutemaker_event_start_time")
+    agenda,
+    event_start_time = getOption("minutemaker_event_start_time")
 ) {
 
   # Import agenda from file
@@ -624,8 +625,9 @@ format_agenda <- function(
     .x$title <- ifelse(is.null(.x$title), "", .x$title)
     .x$description <- ifelse(is.null(.x$description), "", .x$description)
 
-    stringr::str_glue_data(.x,
-                          "Session: {session};
+    stringr::str_glue_data(
+      .x,
+      "Session: {session};
     Title: {title};
     Description: {description};
     Speakers: {speakers};
@@ -697,13 +699,14 @@ import_transcript_from_file <- function(
       # Normal VTT style
       if (stringr::str_detect(cur_speaker, "^\\d+ ")) {
         # the name is between double quotes
-        stringr::str_extract_all('(?<=").*(?=")') |>
-        unlist()
+        cur_speaker <- stringr::str_extract_all(
+          cur_speaker, '(?<=").*(?=")') |>
+          unlist()
       } else if (stringr::str_detect(lines[.x + 1], "^<v ")) {
         # MS Teams vtt style: <v speaker first name [second name]>
         cur_speaker <- stringr::str_extract_all(
           lines[.x + 1], '(?<=<v ).*?(?=>)') |>
-        unlist()
+          unlist()
       } else {
         cur_speaker <- NA
       }
@@ -1097,7 +1100,7 @@ add_chat_transcript <- function(
 #' @param llm_provider A string indicating the LLM provider to use for the
 #'   summarization. See `summarise_transcript` for more details.
 #' @param extra_summarise_args Additional arguments passed to the
-#'   `interrogate_llm` function. See `summarise_transcript` for more details.
+#'   `prompt_llm` function. See `summarise_transcript` for more details.
 #' @param summarization_window_size The size of the summarization window in
 #'   minutes if the "rolling"  method is used. See `summarise_transcript` for
 #'   more details.
@@ -1381,8 +1384,8 @@ speech_to_summary_workflow <- function(
         expected_agenda = expected_agenda,
         window_size = agenda_generation_window_size,
         output_file = if (!purrr::is_empty(agenda) && is.character(agenda)) {
-                          file.path(target_dir, basename(agenda))
-          } else file.path(target_dir, "agenda.R"),
+          file.path(target_dir, basename(agenda))
+        } else file.path(target_dir, "agenda.R"),
         provider = llm_provider
       ), extra_agenda_generation_args)