feat: remove pandoc and its respective engine dependencies

surfsense_backend/app/agents/new_chat/system_prompt.py

@@ -94,6 +94,8 @@ You have access to the following tools:
 
 3. generate_report: Generate a structured Markdown report from provided content.
   - Use this when the user asks to create, generate, write, produce, draft, or summarize into a report-style deliverable.
+  - DECISION RULE (HIGH PRIORITY): If the user asks for a report in any form, call `generate_report` instead of writing the full report directly in chat.
+  - Only skip `generate_report` if the user explicitly asks for chat-only output (e.g., "just answer in chat", "no report card", "don't generate a report").
   - Trigger classes include:
     * Direct trigger words: report, document, memo, letter, template
     * Creation-intent phrases: "write a document/report/post/article"
@@ -108,6 +110,7 @@ You have access to the following tools:
     * "write a report/document", "draft a report"
     * "create an executive summary", "make a briefing note", "write a one-pager"
     * "write a blog post", "write an article", "create a comprehensive guide"
+    * "create a small report", "write a short report", "make a quick report", "brief report for class"
   - IMPORTANT FORMAT RULE: Reports are ALWAYS generated in Markdown.
   - Args:
     - topic: The main topic or title of the report
@@ -121,7 +124,9 @@ You have access to the following tools:
   - Returns: A dictionary with status "ready" or "failed", report_id, title, and word_count.
   - The report is generated immediately in Markdown and displayed inline in the chat.
   - Export/download formats (e.g., PDF/DOCX) are produced from the generated Markdown report.
-  - IMPORTANT: Always search the knowledge base first to gather comprehensive source_content before generating a report.
+  - SOURCE-COLLECTION RULE:
+    * If the user already provided enough source material (current chat content, uploaded files, pasted text, or a summarized video/article), generate the report directly from that.
+    * Use search_knowledge_base first when additional context is needed or the user asks for information beyond what is already available in the conversation.
   - AFTER CALLING THIS TOOL: Do NOT repeat, summarize, or reproduce the report content in the chat. The report is already displayed as an interactive card that the user can open, read, copy, and export. Simply confirm that the report was generated (e.g., "I've generated your report on [topic]. You can view the Markdown report now, and export to PDF/DOCX from the card."). NEVER write out the report text in the chat.
 
 4. link_preview: Fetch metadata for a URL to display a rich preview card.

surfsense_backend/app/agents/new_chat/tools/report.py

@@ -58,6 +58,23 @@ Write the report now:
 """
 
 
+def _strip_wrapping_code_fences(text: str) -> str:
+    """Remove wrapping code fences that LLMs often add around Markdown output.
+
+    Handles patterns like:
+        ```markdown\\n...content...\\n```
+        ```md\\n...content...\\n```
+        ```\\n...content...\\n```
+    """
+    stripped = text.strip()
+    # Match opening fence with optional language tag (markdown, md, or bare)
+    m = re.match(r"^```(?:markdown|md)?\s*\n", stripped)
+    if m and stripped.endswith("```"):
+        stripped = stripped[m.end() :]  # remove opening fence
+        stripped = stripped[:-3].rstrip()  # remove closing fence
+    return stripped
+
+
 def _extract_metadata(content: str) -> dict[str, Any]:
     """Extract metadata from generated Markdown content."""
     # Count section headings
@@ -110,6 +127,11 @@ def create_generate_report_tool(
 
         Use this tool when the user asks to create, generate, write, produce, draft,
         or summarize into a report-style deliverable.
+        HIGH-PRIORITY DECISION RULE:
+        - If the user asks for a report in any form,
+          call this tool rather than writing the full report directly in chat.
+        - Only skip this tool when the user explicitly requests chat-only output and
+          says they do not want a generated report card.
         Trigger classes include:
         - Direct trigger words: report, document, memo, letter, template
         - Creation-intent phrases: "write a document/report/post/article"
@@ -136,11 +158,21 @@ def create_generate_report_tool(
         - "Write an article"
         - "Create a comprehensive guide"
         - "Prepare a report"
+        - "Create a small report"
+        - "Write a short report"
+        - "Make a quick report"
+        - "Brief report for class"
 
         FORMAT/EXPORT RULE:
         - Always generate the report content in Markdown.
         - If the user requests DOCX/Word/PDF or another file format, export from
           the generated Markdown report.
+        SOURCE-COLLECTION RULE:
+        - If enough source material is already present in the conversation (chat
+          history, pasted text, uploaded files, or a provided video/article summary),
+          generate directly from that source_content.
+        - Use knowledge-base search first only when extra context is needed beyond
+          what the user already provided.
 
         VERSIONING — parent_report_id:
         - Set parent_report_id when the user wants to MODIFY, REVISE, IMPROVE,
@@ -298,6 +330,20 @@ def create_generate_report_tool(
                     "title": topic,
                 }
 
+            # LLMs often wrap output in ```markdown ... ``` fences — strip them
+            # so the stored content is clean Markdown.
+            report_content = _strip_wrapping_code_fences(report_content)
+
+            if not report_content:
+                error_msg = "LLM returned empty or invalid content"
+                report_id = await _save_failed_report(error_msg)
+                return {
+                    "status": "failed",
+                    "error": error_msg,
+                    "report_id": report_id,
+                    "title": topic,
+                }
+
             # Extract metadata (includes "status": "ready")
             metadata = _extract_metadata(report_content)