Rewrite all examples using standardization

.pipelines/stages/jobs/custom-nuget-packaging-job.yml

@@ -157,4 +157,3 @@ jobs:
     inputs:
       targetPath: '$(Build.ArtifactStagingDirectory)\nuget'
       artifactName: $(genai_nuget_package_name)
-

examples/c/CMakeLists.txt

@@ -3,9 +3,18 @@ cmake_minimum_required(VERSION 3.18.1)
 project(ortgenaiapp)
 set(CMAKE_CXX_STANDARD 20)
 
+# Download and make available nlohmann/json
+include(FetchContent)
+FetchContent_Declare(
+  nlohmann_json
+  GIT_REPOSITORY https://github.com/nlohmann/json.git
+  GIT_TAG        v3.12.0     # Or update to latest release
+)
+FetchContent_MakeAvailable(nlohmann_json)
+
 option(USE_CXX "Invoke the C++ example" ON)
 option(MODEL_CHAT "Build the Model Chat example" OFF)
-option(MODEL_QA "Build the Model Q&A example without multi-turn prompting" OFF)
+option(MODEL_QA "Build the Model Q&A example" OFF)
 option(MODEL_VISION "Build the Model Vision example" OFF)
 option(PHI4-MM "Build the Phi-4 mm example" OFF)
 option(WHISPER "Build the Whisper example" OFF)
@@ -60,24 +69,29 @@ set(EXAMPLES_SOURCE_DIR ${CMAKE_SOURCE_DIR}/src)
 if(MODEL_CHAT)
   add_executable(model_chat ${EXAMPLES_SOURCE_DIR}/model_chat.cpp ${EXAMPLES_SOURCE_DIR}/common.cpp)
   prepare_executable(model_chat)
+  target_link_libraries(model_chat PRIVATE nlohmann_json::nlohmann_json)
 endif()
 
 if(MODEL_QA)
   add_executable(model_qa ${EXAMPLES_SOURCE_DIR}/model_qa.cpp ${EXAMPLES_SOURCE_DIR}/common.cpp)
   prepare_executable(model_qa)
+  target_link_libraries(model_qa PRIVATE nlohmann_json::nlohmann_json)
 endif()
 
 if(MODEL_VISION)
   add_executable(model_vision ${EXAMPLES_SOURCE_DIR}/model_vision.cpp ${EXAMPLES_SOURCE_DIR}/common.cpp)
   prepare_executable(model_vision)
+  target_link_libraries(model_vision PRIVATE nlohmann_json::nlohmann_json)
 endif()
 
 if(PHI4-MM)
   add_executable(phi4-mm ${CMAKE_SOURCE_DIR}/src/phi4-mm.cpp ${EXAMPLES_SOURCE_DIR}/common.cpp)
   prepare_executable(phi4-mm)
+  target_link_libraries(phi4-mm PRIVATE nlohmann_json::nlohmann_json)
 endif()
 
 if(WHISPER)
   add_executable(whisper ${CMAKE_SOURCE_DIR}/src/whisper.cpp ${EXAMPLES_SOURCE_DIR}/common.cpp)
   prepare_executable(whisper)
+  target_link_libraries(whisper PRIVATE nlohmann_json::nlohmann_json)
 endif()

examples/c/README.md

@@ -1,7 +1,6 @@
 # ONNX Runtime GenAI C Example
 
-> 📝 **Note:** The examples from the main branch of this repository are compatible with the binaries build from the same commit. Therefore, if using the example from `main`, ONNX Runtime GenAI needs to be built from source. If this is your scenario, just build the library and the examples will be auto built along with the library.
-If this is not your scenario, please use prebuilt binaries from the release you're interested in and use the examples from the same version tag and follow the steps below.
+> 📝 **Note:** The examples from the main branch of this repository are compatible with the binaries built from the same commit. Therefore, if using the example from `main`, ONNX Runtime GenAI needs to be built from source. If this is your scenario, just build the library and the examples will be auto built along with the library. If this is not your scenario, please use prebuilt binaries from the release you're interested in and use the examples from the same version tag and follow the steps below.
 
 ## Download the model
 
@@ -67,3 +66,26 @@ For other options to download models, read through [our download options](https:
 2. On Linux and macOS:
    - cd build
    - ./model_qa <path/to/model/from/above> <execution_provider>
+
+## Use constrained decoding
+
+Constrained Decoding is useful when using function/tool calling as it helps in ensuring the output is in the correct format.
+
+We have integrated [LLGuidance](https://github.com/guidance-ai/llguidance) for constrained decoding. There are three types of constrained decoding enabled right now:
+1. Lark Grammar (Recommended): This option allows you to have an option for a regular output as well as function/tool output in JSON format.
+2. JSON Schema: Output will be JSON schema and it will be one of the function/tools provided.
+3. Regex: If a particular regular expression is desired.
+
+To ensure that the function/tool call works correctly with constrained decoding, you need to modify your tokenizer.json file. For each model that has its own tool calling token, the tool calling token's `special` attribute needs to be set to true. For example, Phi-4 mini uses the <|tool_call|> and <|/tool_call|> tokens so you should set the `special` attribute for them as `true` inside `tokenizer.json`.
+
+To run the C examples with function/tool calling:
+```
+# Using JSON Schema with only tool call output
+.\model_chat -m={path to model folder} -e={execution provider} --response_format=json_schema --tools_file={path to json file} --tool_output --tool_call_start="{starting tool call token}" --tool_call_end="{ending tool call token}"
+
+# Using Lark Grammar with only tool call output
+.\model_chat -m={path to model folder} -e={execution provider} --response_format=lark_grammar --tools_file={path to json file} --tool_output --tool_call_start="{starting tool call token}" --tool_call_end="{ending tool call token}"
+
+# Using Lark Grammar with text or tool call output
+.\model_chat -m={path to model folder} -e={execution provider} --response_format=lark_grammar --tools_file={path to json file} --text_output --tool_output --tool_call_start="{starting tool call token}" --tool_call_end="{ending tool call token}"
+```