Merge pull request #519 from presenton/fix/web-search

Fix/web search

electron/servers/fastapi/constants/llm.py

@@ -4,4 +4,4 @@ OPENAI_URL = "https://api.openai.com/v1"
 DEFAULT_OPENAI_MODEL = "gpt-4.1"
 DEFAULT_GOOGLE_MODEL = "models/gemini-2.5-flash"
 DEFAULT_ANTHROPIC_MODEL = "claude-sonnet-4-20250514"
-DEFAULT_CODEX_MODEL = "gpt-5.2-codex"
+DEFAULT_CODEX_MODEL = "gpt-5.1-codex-mini"

electron/servers/fastapi/services/llm_client.py

@@ -100,6 +100,55 @@ class LLMClient:
             return False
         return parse_bool_or_none(get_web_grounding_env()) or False
 
+    def web_search_enabled_for_request(self, web_search: bool) -> bool:
+        """Attach SearchWebTool only when the user enabled web search for this request.
+
+        Controlled solely by the presentation ``web_search`` flag (Advanced settings).
+        Legacy ``WEB_GROUNDING`` / settings toggles are not consulted here so a saved
+        false there cannot disable per-deck web search.
+        """
+        if not web_search:
+            return False
+        if self.llm_provider in (LLMProvider.OLLAMA, LLMProvider.CUSTOM):
+            return False
+        return True
+
+    def outline_uses_prefetched_web_facts(self, web_search: bool) -> bool:
+        """Chat Completions + json_schema rarely invoke custom function tools.
+
+        For OpenAI and Codex we prefetch via the Responses API (``web_search_preview``)
+        and attach the result as context so Advanced settings **Web search** still
+        grounds outlines without relying on ``SearchWebTool`` in the same call.
+        """
+        if not self.web_search_enabled_for_request(web_search):
+            return False
+        return self.llm_provider in (LLMProvider.OPENAI, LLMProvider.CODEX)
+
+    async def prefetch_outline_web_facts(
+        self,
+        content: str,
+        additional_context: Optional[str] = None,
+    ) -> Optional[str]:
+        if self.llm_provider not in (LLMProvider.OPENAI, LLMProvider.CODEX):
+            return None
+        parts = [(content or "").strip(), (additional_context or "").strip()]
+        topic = "\n\n".join(p for p in parts if p)
+        if not topic:
+            topic = "general presentation topic"
+        topic = topic[:12000]
+        query = (
+            "Search the web and summarize the most relevant current facts, statistics, "
+            "and notable recent developments for this presentation topic. Use concise "
+            "bullet points; include approximate dates or time ranges when known.\n\n"
+            f"Topic:\n{topic}"
+        )
+        try:
+            text = await self._search_openai(query)
+            out = (text or "").strip()
+            return out or None
+        except Exception:
+            return None
+
     # ? Disable thinking
     def disable_thinking(self) -> bool:
         return parse_bool_or_none(get_disable_thinking_env()) or False
@@ -1753,7 +1802,7 @@ class LLMClient:
         current_arguments = None
 
         has_response_schema_tool_call = False
-        async for event in await client.chat.completions.create(
+        completion_kwargs: Dict[str, Any] = dict(
             model=model,
             messages=[message.model_dump() for message in messages],
             max_completion_tokens=max_tokens,
@@ -1774,7 +1823,11 @@ class LLMClient:
             ),
             extra_body=extra_body,
             stream=True,
-        ):
+        )
+        if all_tools:
+            completion_kwargs["tool_choice"] = "auto"
+            completion_kwargs["parallel_tool_calls"] = True
+        async for event in await client.chat.completions.create(**completion_kwargs):
             event: OpenAIChatCompletionChunk = event
             if not event.choices:
                 continue

electron/servers/fastapi/services/llm_tool_calls_handler.py

@@ -55,7 +55,7 @@ class LLMToolCallsHandler:
             self.dynamic_tools.append(tool)
 
         match self.client.llm_provider:
-            case LLMProvider.OPENAI | LLMProvider.OLLAMA | LLMProvider.CUSTOM:
+            case LLMProvider.OPENAI | LLMProvider.OLLAMA | LLMProvider.CUSTOM | LLMProvider.CODEX:
                 return self.parse_tool_openai(tool, strict)
             case LLMProvider.ANTHROPIC:
                 return self.parse_tool_anthropic(tool)
@@ -63,7 +63,7 @@ class LLMToolCallsHandler:
                 return self.parse_tool_google(tool)
             case _:
                 raise ValueError(
-                    f"LLM provider must be either openai, anthropic, or google"
+                    "LLM provider must be one of: openai, anthropic, google, codex, ollama, custom"
                 )
 
     def parse_tool_openai(
@@ -181,7 +181,7 @@ class LLMToolCallsHandler:
     # Search web tool call handler
     async def search_web_tool_call_handler(self, arguments: str) -> str:
         match self.client.llm_provider:
-            case LLMProvider.OPENAI:
+            case LLMProvider.OPENAI | LLMProvider.CODEX:
                 return await self.search_web_tool_call_handler_openai(arguments)
             case LLMProvider.ANTHROPIC:
                 return await self.search_web_tool_call_handler_anthropic(arguments)

electron/servers/fastapi/utils/llm_calls/generate_presentation_outlines.py

@@ -16,6 +16,7 @@ def get_system_prompt(
     instructions: Optional[str] = None,
     include_title_slide: bool = True,
     include_table_of_contents: bool = False,
+    web_search: bool = False,
 ):
     verbosity_instruction = (
         "Slide content should be abound 20 words but detailed enough to generate a good slide."
@@ -40,6 +41,27 @@ def get_system_prompt(
     )
     toc_block = f"{toc_instruction}\n" if toc_instruction else ""
 
+    if web_search:
+        tools_hint = "Try to use available tools when they improve accuracy.\n"
+        web_block = (
+            "Web search is enabled: use any \"## Web research (current sources)\" section in Context when present, "
+            "and call SearchWebTool when it is available for fresh facts.\n"
+        )
+    else:
+        tools_hint = ""
+        web_block = "Do not use web search for this outline; rely on Content and Context only.\n"
+
+    url_line = (
+        "Only include URLs if they appear in Content, Context, or a \"## Web research (current sources)\" block.\n"
+        if web_search
+        else "Only include URLs if they appear in the provided content/context.\n"
+    )
+    data_line = (
+        "Ground slide data in Content and Context, and in \"## Web research (current sources)\" when that block is present.\n"
+        if web_search
+        else "Make sure data used is strictly from the provided content/context.\n"
+    )
+
     slide_outline_structure = (
         "Each slide content:\n"
         "   - Must have a ## title.\n"
@@ -60,11 +82,13 @@ def get_system_prompt(
         "If 'auto-detect' is used, figure it out from the content/context.\n"
         f"{title_slide_instruction}\n"
         f"{toc_block}"
+        f"{tools_hint}"
+        f"{web_block}"
         f"{slide_outline_structure}\n"
         "Slide content must not contain any presentation branding/styling information.\n"
         "Title slide must only contain title, presenter name, date and overview.\n"
-        "Only include URLs if they appear in the provided content/context.\n"
-        "Make sure data used is strictly from the provided content/context.\n"
+        f"{url_line}"
+        f"{data_line}"
         "Make sure data is consistent across all slides."
     )
 
@@ -124,6 +148,7 @@ def get_messages(
     instructions: Optional[str] = None,
     include_title_slide: bool = True,
     include_table_of_contents: bool = False,
+    web_search: bool = False,
 ):
     return [
         LLMSystemMessage(
@@ -133,6 +158,7 @@ def get_messages(
                 instructions,
                 include_title_slide,
                 include_table_of_contents,
+                web_search=web_search,
             ),
         ),
         LLMUserMessage(
@@ -171,6 +197,21 @@ async def generate_ppt_outline(
 
     client = LLMClient()
 
+    merged_context = additional_context
+    if client.outline_uses_prefetched_web_facts(web_search):
+        facts = await client.prefetch_outline_web_facts(content, additional_context)
+        if facts:
+            merged_context = (
+                f"{(additional_context or '').strip()}\n\n## Web research (current sources)\n{facts}"
+                if (additional_context or "").strip()
+                else f"## Web research (current sources)\n{facts}"
+            )
+
+    use_search_tool = (
+        client.web_search_enabled_for_request(web_search)
+        and not client.outline_uses_prefetched_web_facts(web_search)
+    )
+
     try:
         async for chunk in client.stream_structured(
             model,
@@ -178,20 +219,17 @@ async def generate_ppt_outline(
                 content,
                 n_slides,
                 language,
-                additional_context,
+                merged_context,
                 tone,
                 verbosity,
                 instructions,
                 include_title_slide,
                 include_table_of_contents,
+                web_search=bool(web_search),
             ),
             response_model.model_json_schema(),
             strict=True,
-            tools=(
-                [SearchWebTool]
-                if (client.enable_web_grounding() and web_search)
-                else None
-            ),
+            tools=([SearchWebTool] if use_search_tool else None),
         ):
             yield chunk
     except Exception as e:

electron/servers/nextjs/app/(presentation-generator)/(dashboard)/settings/SettingCodex.tsx

@@ -46,16 +46,16 @@ interface CodexModel {
 }
 
 const CHATGPT_MODELS: CodexModel[] = [
-    { id: "gpt-5.1", name: "GPT-5.1" },
-    { id: "gpt-5.1-codex-max", name: "GPT-5.1 Codex Max" },
-    { id: "gpt-5.2", name: "GPT-5.2" },
-    { id: "gpt-5.2-codex", name: "GPT-5.2 Codex" },
-    { id: "gpt-5.3-codex", name: "GPT-5.3 Codex" },
-    { id: "gpt-5.4-mini", name: "GPT-5.4 Mini" },
     { id: "gpt-5.4", name: "GPT-5.4" },
+    { id: "gpt-5.2-codex", name: "GPT-5.2-Codex" },
+    { id: "gpt-5.1-codex-max", name: "GPT-5.1-Codex-Max" },
+    { id: "gpt-5.4-mini", name: "GPT-5.4-Mini" },
+    { id: "gpt-5.3-codex", name: "GPT-5.3-Codex" },
+    { id: "gpt-5.2", name: "GPT-5.2" },
+    { id: "gpt-5.1-codex-mini", name: "GPT-5.1-Codex-Mini" },
 ];
 
-const DEFAULT_CODEX_MODEL = "gpt-5.4-mini";
+const DEFAULT_CODEX_MODEL = "gpt-5.1-codex-mini";
 
 export default function CodexConfig({
     codexModel,

electron/servers/nextjs/app/(presentation-generator)/upload/components/AdvanceSettings.tsx

@@ -142,7 +142,7 @@ const AdvanceSettings = ({ config, onConfigChange }: ConfigurationSelectsProps)
                                     onCheckedChange={(checked) => setAdvancedDraft((prev) => ({ ...prev, webSearch: checked }))}
                                 />
                             </div>
-                            <p className="text-xs text-gray-600">Allow the model to consult the web for fresher facts.</p>
+                            <p className="text-xs text-gray-600">Allow the model to consult the web for fresher facts. Turn on and click Save—this toggle alone controls web search for this deck.</p>
                         </div>
 
                         {/* Instructions */}

electron/servers/nextjs/app/(presentation-generator)/upload/components/ConfigurationSelects.tsx

@@ -375,7 +375,7 @@ export function ConfigurationSelects({
                                     onCheckedChange={(checked) => setAdvancedDraft((prev) => ({ ...prev, webSearch: checked }))}
                                 />
                             </div>
-                            <p className="text-xs text-gray-600">Allow the model to consult the web for fresher facts.</p>
+                            <p className="text-xs text-gray-600">Allow the model to consult the web for fresher facts. Turn on and click Save—this toggle alone controls web search for this deck.</p>
                         </div>
 
                         {/* Instructions */}

electron/servers/nextjs/components/CodexConfig.tsx

@@ -33,16 +33,16 @@ interface CodexModel {
 }
 
 export const CHATGPT_MODELS: CodexModel[] = [
-  { id: "gpt-5.1", name: "GPT-5.1" },
-  { id: "gpt-5.1-codex-max", name: "GPT-5.1 Codex Max" },
-  { id: "gpt-5.2", name: "GPT-5.2" },
-  { id: "gpt-5.2-codex", name: "GPT-5.2 Codex" },
-  { id: "gpt-5.3-codex", name: "GPT-5.3 Codex" },
-  { id: "gpt-5.4 mini", name: "GPT-5.4 Mini" },
   { id: "gpt-5.4", name: "GPT-5.4" },
+  { id: "gpt-5.2-codex", name: "GPT-5.2-Codex" },
+  { id: "gpt-5.1-codex-max", name: "GPT-5.1-Codex-Max" },
+  { id: "gpt-5.4-mini", name: "GPT-5.4-Mini" },
+  { id: "gpt-5.3-codex", name: "GPT-5.3-Codex" },
+  { id: "gpt-5.2", name: "GPT-5.2" },
+  { id: "gpt-5.1-codex-mini", name: "GPT-5.1-Codex-Mini" },
 ];
 
-export const DEFAULT_CODEX_MODEL = "gpt-5.4-mini";
+export const DEFAULT_CODEX_MODEL = "gpt-5.1-codex-mini";
 
 export default function CodexConfig({
   codexModel,