feat: Revise prompt and tool descriptions for clarity and accuracy

- Updated the system prompt to provide clearer guidance on the use of deck memory and tools, emphasizing the distinction between historical context and current slide data.
- Enhanced tool descriptions in the ChatTools class to clarify their intended use cases, focusing on live slide data and the importance of using tools for current content.
- Improved instructions on handling slide indices and the retrieval of slide information, ensuring users understand the differences between memory and live data.

servers/fastapi/services/chat/prompts.py

@@ -10,19 +10,31 @@ def build_system_prompt(
     chat_memory_context: str,
 ) -> str:
     presentation_block = _trim_block(
-        "Deck memory:",
+        "Deck memory (semantic / long-term: uploaded document text, outline drafts & prompts, stored slide-edit notes; snippets may be partial and can lag the live deck):",
         presentation_memory_context,
     )
     chat_block = _trim_block(
-        "Chat memory:",
+        "Chat memory (earlier messages in this conversation only):",
         chat_memory_context,
     )
     return (
         "You are Presenton's slide assistant. Be concise, accurate, and action-oriented.\n"
-        "Use tools for live slide data; tool results override memory. Treat user slide numbers as 1-based and tool indexes as 0-based.\n"
-        "Use compact reads first: getPresentationOutline, searchSlides, then getSlideAtIndex with includeFullContent=true only before editing.\n"
-        "Before saving, inspect layouts/schema, batch all needed images/icons with generateAssets, then call saveSlide with schema-valid JSON content.\n"
-        "Do not invent deck facts. When done with tools, stop calling them and answer briefly with what changed or what you found.\n"
+        "\n"
+        "How to use context vs tools\n"
+        "- RAG / deck memory (the block above labeled Deck memory) is the right place to ground answers about: "
+        "the uploaded file or its themes, the outline that was generated, planning intent, and source material. "
+        "It is **not** authoritative for the exact text or order of slides as they exist **right now**.\n"
+        "- For anything about **current** slides (what a slide actually says, layout, which slide is which, or edits to apply), "
+        "use tools. After tools run, their results **override** conflicting deck memory.\n"
+        "- If the user asks about **documents or outlines** (e.g. \"what was in the PDF\", \"original outline\"), search deck memory first; "
+        "if they ask about **what is on slide N** or to **change a slide**, use getPresentationOutline / searchSlides / getSlideAtIndex, not memory alone.\n"
+        "\n"
+        "Tool use (live SQL slide data)\n"
+        "Treat user slide numbers as 1-based; tool slide indexes are 0-based. "
+        "Use compact reads first: getPresentationOutline, searchSlides, then getSlideAtIndex; "
+        "set includeFullContent=true only when you need full JSON (usually right before saveSlide). "
+        "Before saving, inspect layout/schema, batch images/icons with generateAssets, then saveSlide with valid JSON. "
+        "Do not invent deck facts. When finished with tools, stop calling them and answer briefly with what changed or what you found.\n"
         f"{presentation_block}"
         f"{chat_block}"
     )

servers/fastapi/services/chat/tools.py

@@ -43,10 +43,10 @@ class ChatTools:
             Tool(
                 name="getPresentationOutline",
                 description=(
-                    "Retrieve the current presentation outline from live slides "
-                    "database state (with memory fallback when needed). "
-                    "Return compact sections (no full slide JSON) to save context "
-                    "window. Use when the user asks about sections, flow, or slide plan."
+                    "Live database: current deck structure. "
+                    "Use for the **actual** slide list/order and compact previews—not for uploaded PDF text or pre-outline RAG. "
+                    "Falls back to stored outlines only if no slide rows exist. "
+                    "Return compact sections (no full slide JSON). Use for flow, sections, or 'what slides exist'."
                 ),
                 schema=NoArgsInput,
                 strict=True,
@@ -54,8 +54,9 @@ class ChatTools:
             Tool(
                 name="searchSlides",
                 description=(
-                    "Search SQL slides by semantic intent/keywords and return "
-                    "compact relevant snippets with slide identifiers. "
+                    "Live SQL slides: keyword/semantic style search with snippets and indices. "
+                    "Use to find on-slide text, topics, or which slide mentioned something. "
+                    "For source-document-only questions, rely on deck memory; use this when the question is about **slides as built**. "
                     "Always provide both query and limit."
                 ),
                 schema=SearchSlidesInput,
@@ -64,11 +65,9 @@ class ChatTools:
             Tool(
                 name="getSlideAtIndex",
                 description=(
-                    "Retrieve a single slide by zero-based index, including its "
-                    "layout id and compact preview by default. "
-                    "Set includeFullContent=true only when full JSON is explicitly needed "
-                    "(for example before editing existing content). "
-                    "If user says slide N, convert to zero-based index N-1."
+                    "Live SQL: one slide by index—authoritative for exact current content. "
+                    "Set includeFullContent=true when you need full JSON (before saveSlide or precise edits). "
+                    "If user says slide N, use zero-based index N-1."
                 ),
                 schema=GetSlideAtIndexInput,
                 strict=True,