Add DeepSeek support

Fixes #242

DESCRIPTION

@@ -71,6 +71,7 @@ Collate:
     'provider-claude.R'
     'provider-cortex.R'
     'provider-databricks.R'
+    'provider-deepseek.R'
     'provider-gemini.R'
     'provider-github.R'
     'provider-groq.R'

NAMESPACE

@@ -19,6 +19,7 @@ export(chat_bedrock)
 export(chat_claude)
 export(chat_cortex)
 export(chat_databricks)
+export(chat_deepseek)
 export(chat_gemini)
 export(chat_github)
 export(chat_groq)

NEWS.md

@@ -1,5 +1,7 @@
 # ellmer (development version)
 
+* `chat_deepseek()` provides support for DeepSeek models (#242)
+
 * `print(Chat)` no longer wraps long lines, making it easier to read code and bulleted lists (#246).
 
 * `chat_openai()` should be less likely to timeout when not streaming chat results (#213).

R/provider-deepseek.R

@@ -0,0 +1,88 @@
+#' @include provider-openai.R
+NULL
+
+#' Chat with a model hosted on DeepSeek
+#'
+#' @description
+#' Sign up at <https://platform.deepseek.com>.
+#'
+#' This function is a lightweight wrapper around [chat_openai()] with
+#' default settings tailored for DeepSeek.
+#'
+#' ## Known limitations
+#'
+#' * Structured data extraction is not supported..
+#' * Function calling is currently [unstable](https://api-docs.deepseek.com/guides/function_calling).
+#' * Images are not supported.
+#'
+#' @export
+#' @family chatbots
+#' @inheritParams chat_openai
+#' @inherit chat_openai return
+#' @examples
+#' \dontrun{
+#' chat <- chat_deepseek()
+#' chat$chat("Tell me three jokes about statisticians")
+#' }
+chat_deepseek <- function(system_prompt = NULL,
+                          turns = NULL,
+                          api_key = deepseek_key(),
+                          model = NULL,
+                          seed = NULL,
+                          api_args = list(),
+                          echo = NULL) {
+
+  turns <- normalize_turns(turns, system_prompt)
+  model <- set_default(model, "deepseek-chat")
+  echo <- check_echo(echo)
+
+  if (is_testing() && is.null(seed)) {
+    seed <- seed %||% 1014
+  }
+
+  provider <- ProviderDeepSeek(
+    base_url = "https://api.deepseek.com",
+    model = model,
+    seed = seed,
+    extra_args = api_args,
+    api_key = api_key
+  )
+  Chat$new(provider = provider, turns = turns, echo = echo)
+}
+
+ProviderDeepSeek <- new_class("ProviderDeepSeek", parent = ProviderOpenAI)
+
+method(as_json, list(ProviderDeepSeek, ContentText)) <- function(provider, x) {
+  x@text
+}
+
+method(as_json, list(ProviderDeepSeek, Turn)) <- function(provider, x) {
+  if (x@role == "user") {
+    # Text and tool results go in separate messages
+    texts <- keep(x@contents, S7_inherits, ContentText)
+    texts_out <- lapply(texts, function(text) {
+      list(role = "user", content = as_json(provider, text))
+    })
+
+    tools <- keep(x@contents, S7_inherits, ContentToolResult)
+    tools_out <- lapply(tools, function(tool) {
+      list(role = "tool", content = tool_string(tool), tool_call_id = tool@id)
+    })
+
+    c(texts_out, tools_out)
+  } else if (x@role == "assistant") {
+    # Tool requests come out of content and go into own argument
+    text <- detect(x@contents, S7_inherits, ContentText)
+    tools <- keep(x@contents, S7_inherits, ContentToolRequest)
+
+    list(compact(list(
+      role = "assistant",
+      content = as_json(provider, text),
+      tool_calls = as_json(provider, tools)
+    )))
+  } else {
+    as_json(super(provider, ProviderOpenAI), x)
+  }
+}
+
+deepseek_key <- function() key_get("DEEPSEEK_API_KEY")

R/provider-openai.R

@@ -116,6 +116,8 @@ method(chat_request, ProviderOpenAI) <- function(provider,
   req <- req_error(req, body = function(resp) {
     if (resp_content_type(resp) == "application/json") {
       resp_body_json(resp)$error$message
+    } else if (resp_content_type(resp) == "text/plain") {
+      resp_body_string(resp)
     }
   })
 

README.Rmd

@@ -41,6 +41,7 @@ ellmer supports a wide variety of model providers:
 * AWS Bedrock: `chat_bedrock()`.
 * Azure OpenAI: `chat_azure()`.
 * Databricks: `chat_databricks()`.
+* DeepSeek: `chat_deepseek()`.
 * GitHub model marketplace: `chat_github()`.
 * Google Gemini: `chat_gemini()`.
 * Groq: `chat_groq()`.

README.md

@@ -34,6 +34,7 @@ ellmer supports a wide variety of model providers:
 - AWS Bedrock: `chat_bedrock()`.
 - Azure OpenAI: `chat_azure()`.
 - Databricks: `chat_databricks()`.
+- DeepSeek: `chat_deepseek()`.
 - GitHub model marketplace: `chat_github()`.
 - Google Gemini: `chat_gemini()`.
 - Groq: `chat_groq()`.

tests/testthat/_snaps/provider-deepseek.md

@@ -0,0 +1,7 @@
+# defaults are reported
+
+    Code
+      . <- chat_deepseek()
+    Message
+      Using model = "deepseek-chat".
+

tests/testthat/test-provider-deepseek.R

@@ -0,0 +1,52 @@
+# Getting started --------------------------------------------------------
+
+test_that("can make simple request", {
+  chat <- chat_deepseek("Be as terse as possible; no punctuation")
+  resp <- chat$chat("What is 1 + 1?", echo = FALSE)
+  expect_match(resp, "2")
+  expect_equal(chat$last_turn()@tokens, c(20, 1))
+})
+
+test_that("can make simple streaming request", {
+  chat <- chat_deepseek("Be as terse as possible; no punctuation")
+  resp <- coro::collect(chat$stream("What is 1 + 1?"))
+  expect_match(paste0(unlist(resp), collapse = ""), "2")
+})
+
+# Common provider interface -----------------------------------------------
+
+test_that("defaults are reported", {
+  expect_snapshot(. <- chat_deepseek())
+})
+
+test_that("respects turns interface", {
+  chat_fun <- chat_deepseek
+
+  test_turns_system(chat_fun)
+  test_turns_existing(chat_fun)
+})
+
+# Only partially works
+# test_that("all tool variations work", {
+#   chat_fun <- chat_deepseek
+
+#   test_tools_simple(chat_fun)
+#   test_tools_async(chat_fun)
+#   test_tools_parallel(chat_fun)
+#   test_tools_sequential(chat_fun, total_calls = 6)
+# })
+
+# # Doesn't support data extraction
+# test_that("can extract data", {
+#   chat_fun <- chat_deepseek
+
+#   test_data_extraction(chat_fun)
+# })
+
+# # Doesn't support images
+# test_that("can use images", {
+#   chat_fun <- chat_deepseek_test
+
+#   test_images_inline(chat_fun)
+#   test_images_remote(chat_fun)
+# })