Merge pull request #3 from ChatFAQ/develop

sync with ChatFAQ

Author: Hectruendo

admin/components/labeling/KnowledgeItemReview.vue

@@ -76,8 +76,12 @@ async function initKIReview() {
     }
     for (const ki_ref of props.references.knowledge_items || []) {
         const ki = await itemsStore.retrieveItems("/back/api/language-model/knowledge-items/", {id: ki_ref.knowledge_item_id, limit: 0, offset: 0, ordering: undefined}, true)
-        if (ki)
+        if (ki) {
             reviewedKIs.value.kis.push(ki)
+        } else if (ki_ref.title && ki_ref.content) {
+            // If KI not found in backend, use the reference if it has title and content
+            reviewedKIs.value.kis.push(ki_ref);
+        }
     }
     review.value = await itemsStore.retrieveItems("/back/api/broker/admin-review/", {message: props.message.id, limit: 0, offset: 0, ordering: undefined}, true) || {}
     ki_choices.value = (await itemsStore.retrieveItems("/back/api/language-model/knowledge-items/", {knowledge_base: props.references.knowledge_base_id, knowledge_base__id: props.references.knowledge_base_id, limit: 0, offset: 0, ordering: undefined})).results

chat_rag/chat_rag/llms/format_tools.py

@@ -92,7 +92,7 @@ def gemini_schema(schema: Dict) -> Dict[str, Any]:
     schema = {
         "name": schema["name"],
         "description": schema["description"],
-        "parameters": uppercase_types_recursively(schema["parameters"]),
+        "parameters": transform_tool_spec(uppercase_types_recursively(schema["parameters"]), keys_to_remove=["additionalProperties", "default", "$schema"]),
     }
     return schema
 
@@ -113,6 +113,22 @@ def uppercase_types_recursively(schema: Dict[str, Any]) -> Dict[str, Any]:
             schema[i] = uppercase_types_recursively(item)
     return schema
 
+def transform_tool_spec(data, keys_to_remove):
+    """
+    Recursively removes 'additionalProperties', 'default', and '$schema' keys
+    from a dictionary or list representing tool specifications.
+    """
+    if isinstance(data, dict):
+        new_dict = {}
+        for key, value in data.items():
+            if key not in keys_to_remove:
+                new_dict[key] = transform_tool_spec(value, keys_to_remove)
+        return new_dict
+    elif isinstance(data, list):
+        return [transform_tool_spec(item, keys_to_remove) for item in data]
+    else:
+        return data
+
 
 def _check_tool_choice(tools: List[Dict], tool_choice: str) -> str:
     """
@@ -154,6 +170,7 @@ def format_tools(
     tools_formatted = []
     if mode in {Mode.OPENAI_TOOLS, Mode.MISTRAL_TOOLS}:
         for tool in tools:
+            tool = transform_tool_spec(tool, keys_to_remove=["default"])
             # As it is already in the openai format, we can just append it
             tools_formatted.append(tool)
 

chat_rag/chat_rag/llms/mistral_client.py

@@ -3,7 +3,6 @@
 
 from mistralai.async_client import MistralAsyncClient
 from mistralai.client import MistralClient
-from mistralai.models.chat_completion import ChatMessage
 
 from .base_llm import LLM
 from .format_tools import Mode, format_tools
@@ -21,12 +20,12 @@ def __init__(
     def format_prompt(
         self,
         messages: List[Dict[str, str]],
-    ) -> List[ChatMessage]:
+    ) -> List[Dict[str, str]]:
         """
         Formats the prompt to be used by the model into the correct Mistral format.
         """
         final_messages = [
-            ChatMessage(role=message["role"], content=message["content"])
+            {"role": message["role"], "content": message["content"]}
             for message in messages
         ]
 

chat_rag/chat_rag/llms/openai_client.py

@@ -166,7 +166,7 @@ def stream(
             seed=seed,
             n=1,
             stream=True,
-            reasoning_effort=thinking,
+            reasoning_effort=thinking if thinking else NOT_GIVEN,
         )
         for chunk in response:
             if chunk.choices[0].finish_reason == "stop":
@@ -203,7 +203,7 @@ async def astream(
             seed=seed,
             n=1,
             stream=True,
-            reasoning_effort=thinking,
+            reasoning_effort=thinking if thinking else NOT_GIVEN,
         )
         async for chunk in response:
             if chunk.choices[0].finish_reason == "stop":
@@ -244,7 +244,7 @@ def generate(
             temperature=temperature,
             max_completion_tokens=max_tokens,
             seed=seed,
-            reasoning_effort=thinking,
+            reasoning_effort=thinking if thinking else NOT_GIVEN,
             n=1,
             tools=tools,
             tool_choice=tool_choice,
@@ -285,7 +285,7 @@ async def agenerate(
             temperature=temperature,
             max_completion_tokens=max_tokens,
             seed=seed,
-            reasoning_effort=thinking,
+            reasoning_effort=thinking if thinking else NOT_GIVEN,
             n=1,
             tools=tools,
             tool_choice=tool_choice,

chat_rag/poetry.lock

@@ -19,7 +19,7 @@ aiohttp = "^3.8.5"
 cryptography = "^41.0.4"
 openai = "1.66.2"
 anthropic = "0.49.0"
-mistralai = "0.4.0"
+mistralai = "1.6.0"
 docstring-parser = "^0.16"
 torch = {version = "2.3.0", optional = true}
 transformers = {version = "^4.41.2", optional = true}
@@ -34,7 +34,7 @@ ragatouille = {version = "^0.0.8.post2", optional = true}
 hdbscan = {version = "^0.8.36", optional = true}
 umap-learn = {version = "^0.5.6", optional = true}
 bm25s = {version = "^0.1.7", optional = true}
-google-genai = "1.2.0"
+google-genai = "1.9.0"
 pillow = "^11.1.0"
 together = "^1.4.1"