branch: externals/llm
commit f9e38f6c75506296911f1044b8fe8d41a55499e6
Author: Andrew Hyatt <ahy...@gmail.com>
Commit: GitHub <nore...@github.com>
Add Gemini 2.5 reasoning support, and flash lite model (#202)
Also change the default Vertex model to Gemini 2.5 Pro.
This fixes https://github.com/ahyatt/llm/issues/187.
---
NEWS.org | 4 ++++
llm-gemini.el | 9 +++++----
llm-models.el | 12 +++++++++---
llm-test.el | 5 +++--
llm-vertex.el | 38 ++++++++++++++++++++++++++++++++------
5 files changed, 53 insertions(+), 15 deletions(-)
diff --git a/NEWS.org b/NEWS.org
index 5ad76a1d7f..56a60b214b 100644
--- a/NEWS.org
+++ b/NEWS.org
@@ -1,3 +1,7 @@
+* Version 0.27.1
+- Add thinking control to Gemini / Vertex.
+- Change default Vertex model to Gemini 2.5 Pro.
+- Add Gemini 2.5 Flash model
* Version 0.27.0
- Add =thinking= option to control the amount of thinking that happens for
reasoning models.
- Fix incorrectly low default Claude max tokens
diff --git a/llm-gemini.el b/llm-gemini.el
index 38dfbca007..bb3fbd7c46 100644
--- a/llm-gemini.el
+++ b/llm-gemini.el
@@ -78,10 +78,11 @@ If STREAMING-P is non-nil, use the streaming endpoint."
(cl-defmethod llm-provider-chat-streaming-url ((provider llm-gemini))
(llm-gemini--chat-url provider t))
-(cl-defmethod llm-provider-chat-request ((_ llm-gemini) _ _)
- ;; Temporary, can be removed in the next version. Without this the old
- ;; definition will cause problems when users upgrade.
- (cl-call-next-method))
+(cl-defmethod llm-provider-chat-request ((provider llm-gemini) prompt _)
+ (llm-provider--chat-request prompt (let ((model (llm-models-match
(llm-gemini-chat-model provider))))
+ (if model
+ (llm-model-symbol model)
+ 'unknown))))
(cl-defmethod llm-name ((_ llm-gemini))
"Return the name of PROVIDER."
diff --git a/llm-models.el b/llm-models.el
index d57823e1d2..8f4cded106 100644
--- a/llm-models.el
+++ b/llm-models.el
@@ -174,15 +174,21 @@ REGEX is a regular expression that can be used to
identify the model, uniquely (
:regex "gemini-2\\.0-flash-thinking")
(make-llm-model
:name "Gemini 2.5 Pro" :symbol 'gemini-2.5-pro
- :capabilities '(generation tool-use image-input audio-input video-input
json-response)
+ :capabilities '(generation tool-use image-input audio-input video-input
json-response reasoning)
:context-length 1048576
:regex "gemini-2\\.5-pro")
+ (make-llm-model
+ :name "Gemini 2.5 Flash Lite" :symbol 'gemini-2.5-flash-lite
+ :capabilities '(generation tool-use image-input audio-input video-input
json-response
+ pdf-input caching reasoning)
+ :context-length 1048576
+ :regex "gemini-2\\.5-flash-lite")
(make-llm-model
:name "Gemini 2.5 Flash" :symbol 'gemini-2.5-flash
:capabilities '(generation tool-use image-input audio-input video-input
json-response
- pdf-input caching)
+ pdf-input caching reasoning)
:context-length 1048576
- :regex "gemini-2\\.5-flash")
+ :regex "gemini-2\\.5-flash$")
(make-llm-model
:name "Gemini 2.0 Pro" :symbol 'gemini-2.0-pro
:capabilities '(generation tool-use image-input audio-input video-input)
diff --git a/llm-test.el b/llm-test.el
index 1d68a5f726..9ae1fc0dc5 100644
--- a/llm-test.el
+++ b/llm-test.el
@@ -319,9 +319,10 @@
(should (member 'tool-use (llm-capabilities (make-llm-gemini :chat-model
"gemini-1.5-flash")))))
(ert-deftest llm-test-chat-token-limit-vertex ()
- (should (= 2097152 (llm-chat-token-limit (make-llm-vertex))))
+ ;; The default is Gemini 2.5 Pro, which has a token limit of 1048576.
+ (should (= 1048576 (llm-chat-token-limit (make-llm-vertex))))
(should (= 1048576 (llm-chat-token-limit
- (make-llm-gemini :chat-model "gemini-1.5-flash"))))
+ (make-llm-gemini :chat-model "gemini-2.5-flash"))))
(should (= 4096 (llm-chat-token-limit
(make-llm-vertex :chat-model "unknown")))))
diff --git a/llm-vertex.el b/llm-vertex.el
index 071e32ac67..852592ad87 100644
--- a/llm-vertex.el
+++ b/llm-vertex.el
@@ -59,7 +59,7 @@ and there is no default. The maximum value possible here is
2049."
:type 'integer
:group 'llm-vertex)
-(defcustom llm-vertex-default-chat-model "gemini-1.5-pro"
+(defcustom llm-vertex-default-chat-model "gemini-2.5-pro"
"The default model to ask for.
This should almost certainly be a chat model, other models are
for more specialized uses."
@@ -207,7 +207,8 @@ the key must be regenerated every hour."
(llm-multipart-parts
(llm-chat-prompt-interaction-content interaction)))))
(t `[(:text ,(llm-chat-prompt-interaction-content
interaction))]))))
-(cl-defmethod llm-provider-chat-request ((_ llm-google) prompt _)
+(defun llm-provider--chat-request (prompt model)
+ "Create the request for the chat PROMPT and MODEL symbol."
(llm-provider-utils-combine-to-system-prompt prompt
llm-vertex-example-prelude)
(append
(when (eq 'system (llm-chat-prompt-interaction-role (car
(llm-chat-prompt-interactions prompt))))
@@ -231,12 +232,25 @@ the key must be regenerated every hour."
:parameters ,(llm-provider-utils-openai-arguments
(llm-tool-args tool))))
(llm-chat-prompt-tools prompt))))]))
- (llm-vertex--chat-parameters prompt)))
+ (llm-vertex--chat-parameters prompt model)))
-(defun llm-vertex--chat-parameters (prompt)
+;; TODO: remove after September 2025, this is only here so people can upgrade
to
+;; a new version of their llm library without the old llm-google specializer
+;; sticking around.
+(cl-defmethod llm-provider-chat-request ((_ llm-google) _ _)
+ (cl-call-next-method))
+
+(cl-defmethod llm-provider-chat-request ((provider llm-vertex) prompt _)
+ (llm-provider--chat-request prompt (let ((model (llm-models-match
(llm-vertex-chat-model provider))))
+ (if model
+ (llm-model-symbol model)
+ 'unknown))))
+
+(defun llm-vertex--chat-parameters (prompt model)
"From PROMPT, create the parameters section.
-Return value is a cons for adding to an alist, unless there is
-nothing to add, in which case it is nil."
+Return value is a cons for adding to an alist, unless there is nothing
+to add, in which case it is nil. MODEL is the symbol of the model used,
+which is necessary to properly set some paremeters."
(let ((params-plist (llm-provider-utils-non-standard-params-plist prompt)))
(when (llm-chat-prompt-temperature prompt)
(setq params-plist (plist-put params-plist :temperature
@@ -251,6 +265,18 @@ nothing to add, in which case it is nil."
(setq params-plist (plist-put params-plist :response_schema
(llm-provider-utils-convert-to-serializable
(llm-chat-prompt-response-format
prompt))))))
+ (when-let ((budget (llm-chat-prompt-reasoning prompt))
+ (max-budget (if (eq model 'gemini-2.5-pro) 32768 24576)))
+ (when (member 'reasoning (llm-model-capabilities (llm-models-by-symbol
model)))
+ (if (and (eq model 'gemini-2.5-pro) (eq budget 'none))
+ (display-warning 'llm :warning "Cannot turn off reasoning in
Gemini 2.5 Pro, ignoring reasoning setting")
+ (setq params-plist (plist-put params-plist :thinkingConfig
+ `(:thinkingBudget
+ ,(pcase budget
+ ('none 0)
+ ('light 1024)
+ ('medium (/ max-budget 2))
+ ('maximum max-budget))))))))
(when params-plist
`(:generationConfig ,params-plist))))
