From 3d7be5888e9fb64f7d761b2ed51154d73f4a8a2a Mon Sep 17 00:00:00 2001
From: Bryan Van Deusen <bryan.vandeusen@fabledsword.com>
Date: Mon, 2 Mar 2026 18:41:49 -0500
Subject: [PATCH] Remove intent model entirely; quick-capture uses primary
 model

The separate intent model (OLLAMA_INTENT_MODEL / qwen2.5:7b) is removed
from every part of the system. All classification now uses the primary model.

Changes:
- config.py: remove OLLAMA_INTENT_MODEL
- intent.py: remove classify_intent() and all supporting infrastructure
  (_SYSTEM_PROMPT_TEMPLATE, _RESEARCH_PREFIX, _PRIOR_WORK_REFS); file now
  only contains the quick-capture classifier
- quick_capture.py: classify_capture_intent() now called with Config.OLLAMA_MODEL
- generation_task.py: remove intent_model_setting DB lookup and get_setting import;
  history summarization and research pipeline use the primary model directly
- research.py: remove intent_model parameter from run_research_pipeline() and
  _generate_sub_queries(); both use the model param throughout
- routes/settings.py: remove intent_model from model-key validation and response
- app.py: remove intent model pre-warming at startup
- SettingsView.vue: remove Intent Model selector and related refs/state

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 frontend/src/views/SettingsView.vue           |  15 +-
 src/fabledassistant/app.py                    |  10 +-
 src/fabledassistant/config.py                 |   3 -
 src/fabledassistant/routes/quick_capture.py   |   9 +-
 src/fabledassistant/routes/settings.py        |  16 +-
 .../services/generation_task.py               |  17 +-
 src/fabledassistant/services/intent.py        | 170 +-----------------
 src/fabledassistant/services/research.py      |   9 +-
 8 files changed, 26 insertions(+), 223 deletions(-)

diff --git a/frontend/src/views/SettingsView.vue b/frontend/src/views/SettingsView.vue
index f55d30c..e27e31c 100644
--- a/frontend/src/views/SettingsView.vue
+++ b/frontend/src/views/SettingsView.vue
@@ -10,10 +10,8 @@ const authStore = useAuthStore();
 const toastStore = useToastStore();
 const assistantName = ref("");
 const defaultModel = ref("");
-const intentModel = ref("");
 const installedModels = ref<string[]>([]);
 const defaultChatModel = ref("");
-const defaultIntentModel = ref("");
 const newEmail = ref("");
 const emailPassword = ref("");
 const changingEmail = ref(false);
@@ -80,10 +78,9 @@ onMounted(async () => {
 
   // Load installed models and configured defaults
   try {
-    const modelsData = await apiGet<{ models: string[]; default_chat_model: string; default_intent_model: string }>("/api/settings/models");
+    const modelsData = await apiGet<{ models: string[]; default_chat_model: string }>("/api/settings/models");
     installedModels.value = modelsData.models;
     defaultChatModel.value = modelsData.default_chat_model;
-    defaultIntentModel.value = modelsData.default_intent_model;
   } catch {
     // Ollama unreachable — dropdowns will be empty
   }
@@ -91,7 +88,6 @@ onMounted(async () => {
   // Load notification preferences from user settings
   const allSettings = await apiGet<Record<string, string>>("/api/settings");
   defaultModel.value = allSettings.default_model ?? "";
-  intentModel.value = allSettings.intent_model ?? "";
   if (allSettings.notify_task_reminders !== undefined) {
     notifyTaskReminders.value = allSettings.notify_task_reminders !== "false";
   }
@@ -190,7 +186,6 @@ async function saveAssistant() {
     await store.updateSettings({
       assistant_name: assistantName.value.trim() || "Fable",
       default_model: defaultModel.value,
-      intent_model: intentModel.value,
     });
     saved.value = true;
     setTimeout(() => (saved.value = false), 2000);
@@ -414,14 +409,6 @@ function hostname(url: string): string {
             </select>
             <p class="field-hint">Model used for new conversations.</p>
           </div>
-          <div class="field">
-            <label for="intent-model">Intent Model</label>
-            <select id="intent-model" v-model="intentModel" class="input">
-              <option value="">Default ({{ defaultIntentModel || "qwen2.5:1.5b" }})</option>
-              <option v-for="m in installedModels" :key="m" :value="m">{{ m }}</option>
-            </select>
-            <p class="field-hint">Smaller/faster model for intent routing before the main model.</p>
-          </div>
         </div>
         <div class="actions">
           <button class="btn-save" @click="saveAssistant" :disabled="saving">
diff --git a/src/fabledassistant/app.py b/src/fabledassistant/app.py
index 39ce4a9..ebeffea 100644
--- a/src/fabledassistant/app.py
+++ b/src/fabledassistant/app.py
@@ -131,15 +131,9 @@ def create_app() -> Quart:
                 except Exception:
                     logger.warning("Failed to warm model '%s'", model, exc_info=True)
 
-        # Pull main model and (if configured) a separate intent model concurrently,
-        # then warm them into VRAM so the first user request doesn't cold-load both
-        # simultaneously (which causes Ollama 500 races).
-        # Fire-and-forget so pulls/warming don't block startup.
+        # Pull and warm the main model into VRAM at startup so the first request is fast.
+        asyncio.create_task(_pull_model(Config.OLLAMA_MODEL, warm=True))
         models_to_warm = {Config.OLLAMA_MODEL}
-        if Config.OLLAMA_INTENT_MODEL and Config.OLLAMA_INTENT_MODEL != Config.OLLAMA_MODEL:
-            models_to_warm.add(Config.OLLAMA_INTENT_MODEL)
-        for _model in models_to_warm:
-            asyncio.create_task(_pull_model(_model, warm=True))
         # Also pull the embedding model (nomic-embed-text by default), but no need to warm it.
         if Config.EMBEDDING_MODEL not in models_to_warm:
             asyncio.create_task(_pull_model(Config.EMBEDDING_MODEL, warm=False))
diff --git a/src/fabledassistant/config.py b/src/fabledassistant/config.py
index 67d9c68..ef9df09 100644
--- a/src/fabledassistant/config.py
+++ b/src/fabledassistant/config.py
@@ -24,9 +24,6 @@ class Config:
     )
     OLLAMA_URL: str = os.environ.get("OLLAMA_URL", "http://localhost:11434")
     OLLAMA_MODEL: str = os.environ.get("OLLAMA_MODEL", "qwen3:latest")
-    # Optional dedicated model for intent classification.
-    # Falls back to OLLAMA_MODEL if not set. Can also be overridden per-user via settings.
-    OLLAMA_INTENT_MODEL: str = os.environ.get("OLLAMA_INTENT_MODEL", "qwen2.5:7b")
     # KV cache context window for generation. Lower = less VRAM, less throughput impact.
     OLLAMA_NUM_CTX: int = int(os.environ.get("OLLAMA_NUM_CTX", "16384"))
     SECRET_KEY: str = _read_secret("SECRET_KEY", "SECRET_KEY_FILE", "dev-secret-change-me")
diff --git a/src/fabledassistant/routes/quick_capture.py b/src/fabledassistant/routes/quick_capture.py
index 37bc136..f980aad 100644
--- a/src/fabledassistant/routes/quick_capture.py
+++ b/src/fabledassistant/routes/quick_capture.py
@@ -78,7 +78,7 @@ async def quick_capture_route():
     if not text:
         return jsonify({"error": "text is required"}), 400
 
-    intent_model = Config.OLLAMA_INTENT_MODEL or Config.OLLAMA_MODEL
+    model = Config.OLLAMA_MODEL
 
     # Build tool list for this user, then restrict to capture-only operations.
     all_tools = await get_tools_for_user(uid)
@@ -86,7 +86,7 @@ async def quick_capture_route():
         t for t in all_tools if t.get("function", {}).get("name") in _CAPTURE_TOOL_NAMES
     ]
 
-    intent = await classify_capture_intent(text, capture_tools, intent_model)
+    intent = await classify_capture_intent(text, capture_tools, model)
 
     if intent.should_execute:
         # research_topic bypasses execute_tool — run the pipeline directly
@@ -94,9 +94,8 @@ async def quick_capture_route():
             from fabledassistant.services.research import run_research_pipeline
 
             topic = intent.arguments.get("topic", text)
-            model = Config.OLLAMA_MODEL
             try:
-                note = await run_research_pipeline(topic, uid, model, intent_model)
+                note = await run_research_pipeline(topic, uid, model)
                 logger.info(
                     "Quick-capture uid=%d: research note id=%d '%s'",
                     uid, note.id, note.title,
@@ -114,7 +113,6 @@ async def quick_capture_route():
         # For notes, run a second LLM pass to generate a proper title and
         # well-formed body rather than using the raw capture text verbatim.
         if intent.tool_name == "create_note":
-            model = Config.OLLAMA_MODEL
             title, body = await _process_note(text, model)
             intent.arguments["title"] = title
             intent.arguments["body"] = body
@@ -141,7 +139,6 @@ async def quick_capture_route():
 
     # Fallback: classify_capture_intent returned no-tool (e.g. LLM parse failure).
     # Still process the text through the note enrichment pass.
-    model = Config.OLLAMA_MODEL
     fallback_title, fallback_body = await _process_note(text, model)
 
     result = await execute_tool(
diff --git a/src/fabledassistant/routes/settings.py b/src/fabledassistant/routes/settings.py
index ebe5e22..e51a471 100644
--- a/src/fabledassistant/routes/settings.py
+++ b/src/fabledassistant/routes/settings.py
@@ -25,18 +25,17 @@ async def update_settings_route():
     if not isinstance(data, dict):
         return jsonify({"error": "Expected a JSON object"}), 400
 
-    if "default_model" in data or "intent_model" in data:
+    if "default_model" in data:
         installed = await get_installed_models()
-        for key in ("default_model", "intent_model"):
-            if key in data and data[key]:
-                model = str(data[key])
-                if installed and model not in installed:
-                    return jsonify({"error": f"Model '{model}' is not installed"}), 400
+        if data["default_model"]:
+            model = str(data["default_model"])
+            if installed and model not in installed:
+                return jsonify({"error": f"Model '{model}' is not installed"}), 400
 
-    # Empty string for model keys means "reset to system default".
+    # Empty string for default_model means "reset to system default".
     # Delete the DB row so get_setting() falls back to Config defaults
     # rather than returning "" and breaking model resolution everywhere.
-    _MODEL_KEYS = frozenset({"default_model", "intent_model"})
+    _MODEL_KEYS = frozenset({"default_model"})
     to_save = {}
     for k, v in data.items():
         str_v = str(v)
@@ -59,7 +58,6 @@ async def get_models_route():
     return jsonify({
         "models": models,
         "default_chat_model": Config.OLLAMA_MODEL,
-        "default_intent_model": Config.OLLAMA_INTENT_MODEL,
     })
 
 
diff --git a/src/fabledassistant/services/generation_task.py b/src/fabledassistant/services/generation_task.py
index 16c730c..7952fee 100644
--- a/src/fabledassistant/services/generation_task.py
+++ b/src/fabledassistant/services/generation_task.py
@@ -25,7 +25,6 @@ from fabledassistant.services.generation_buffer import (
 from fabledassistant.services.llm import ChatChunk, build_context, generate_completion, stream_chat, stream_chat_with_tools, summarize_history_for_context, wait_for_model_loaded
 from fabledassistant.services.chat import update_conversation_title
 from fabledassistant.services.logging import log_generation
-from fabledassistant.services.settings import get_setting
 from fabledassistant.services.tools import get_tools_for_user, execute_tool
 from fabledassistant.services.research import run_research_pipeline
 
@@ -156,16 +155,12 @@ async def run_generation(
 
     buf.append_event("status", {"status": "Building context..."})
 
-    # Phase 1: Quick DB calls — resolve tools list and intent model in parallel.
-    tools, intent_model_setting = await asyncio.gather(
-        get_tools_for_user(user_id),
-        get_setting(user_id, "intent_model", ""),
-    )
-    intent_model = intent_model_setting or Config.OLLAMA_INTENT_MODEL or model
+    # Phase 1: Resolve the tools list for this user.
+    tools = await get_tools_for_user(user_id)
 
     logger.info(
-        "Starting generation for conv %d: model=%s, intent_model=%s, tools=%d",
-        conv_id, model, intent_model, len(tools),
+        "Starting generation for conv %d: model=%s, tools=%d",
+        conv_id, model, len(tools),
     )
 
     # Phase 2: Summarize long conversation history if needed.
@@ -173,7 +168,7 @@ async def run_generation(
     history_summary: str | None = None
     if len(history) > 20:  # matches _HISTORY_SUMMARY_THRESHOLD in llm.py
         buf.append_event("status", {"status": "Summarizing conversation history..."})
-        history_to_use, history_summary = await summarize_history_for_context(history, intent_model)
+        history_to_use, history_summary = await summarize_history_for_context(history, model)
 
     # Phase 3: Build context and wait for model in parallel.
     model_load_task = asyncio.create_task(wait_for_model_loaded(model, timeout=90.0))
@@ -259,7 +254,7 @@ async def run_generation(
                         if tool_name == "research_topic":
                             topic = arguments.get("topic", "")
                             try:
-                                note = await run_research_pipeline(topic, user_id, model, intent_model, buf)
+                                note = await run_research_pipeline(topic, user_id, model, buf)
                                 result = {
                                     "success": True,
                                     "type": "research_note",
diff --git a/src/fabledassistant/services/intent.py b/src/fabledassistant/services/intent.py
index 229804e..13b1fbe 100644
--- a/src/fabledassistant/services/intent.py
+++ b/src/fabledassistant/services/intent.py
@@ -1,9 +1,7 @@
-"""Intent routing — classify user message before streaming.
+"""Quick-capture intent classifier.
 
-Makes a fast non-streaming LLM call to detect tool intent and extract
-parameters.  When a tool call is detected the caller can execute it
-directly, bypassing the model's native (and sometimes unreliable)
-structured tool-calling API.
+Classifies short capture text (note, task, event, research) for the
+/api/quick-capture endpoint using a dedicated prompt and the primary model.
 """
 
 import json
@@ -51,168 +49,6 @@ def _build_tool_summary(tools: list[dict]) -> str:
     return "\n".join(lines)
 
 
-_SYSTEM_PROMPT_TEMPLATE = """\
-You are an intent classifier. Given a user message (and recent conversation \
-history for context), decide whether it requires calling one of the available \
-tools or is just general chat.
-
-Today's date is {today}.
-
-Available tools:
-{tool_summary}
-
-Respond with ONLY a JSON object, no other text:
-- If a tool should be called: {{"tool": "tool_name", "arguments": {{...}}, "confidence": "high"|"medium"|"low", "ack": "One short sentence describing what you're about to do."}}
-- If it's general chat: {{"tool": null, "confidence": "high", "ack": null}}
-
-Confidence levels:
-- "high": the intent is clear and all required arguments are unambiguous
-- "medium": probably requires the tool but some argument is uncertain or inferred
-- "low": uncertain whether this needs a tool at all, or the message is too ambiguous to act on
-
-Rules:
-- Use recent conversation history to resolve references like "it", "that event", "the meeting", "move it", etc.
-- For dates like "tomorrow", "next Friday", "in 3 days", resolve them to YYYY-MM-DD format.
-- For datetime parameters, use ISO 8601 format (e.g. 2026-09-30T14:00:00).
-- Only include arguments the user actually specified or that can be clearly inferred.
-- Infer reasonable defaults: birthdays and holidays are all-day + yearly recurring; "weekly meeting" is weekly recurring.
-- Use descriptive titles: "My Birthday" not just "Birthday", "Team Standup" not just "Meeting".
-- "add to", "update", "edit", "expand", "flesh out", "modify", "append to", "continue writing" a note → use update_note with query=<note title from context> and mode="append" for additions or mode="replace" for full rewrites.
-- "mark as done", "complete", "finish", "mark in progress", "start" a task → use update_note with status field.
-- "set priority", "change priority", "make it high priority" → use update_note with priority field.
-- "set due date", "move due date", "due on Friday" for a task → use update_note with due_date field.
-- "what are my overdue tasks", "show overdue", "tasks due today", "high priority tasks", "in progress tasks", "what's due this week" → use list_tasks with appropriate status/priority/due_before/due_after filters. For overdue, set due_before to today's date.
-- If a note was created earlier in the conversation and the user provides more content for it, use update_note (not create_note).
-- Use create_note ONLY when the user explicitly wants a brand new note that doesn't already exist.
-- "update", "change", "move", "reschedule" an event → use update_event.
-- "delete", "cancel", "remove" an event → use delete_event.
-- "which calendars", "list calendars", "my calendars" → use list_calendars.
-- "remind me X minutes/hours before" an event → convert to reminder_minutes parameter on create_event.
-- When the user asks about events in a time period (e.g. "events in September", "what's on next week", "my schedule for March"), use list_events with date_from/date_to covering that period. Do NOT use search_events for time-based queries — search_events is only for keyword matching against event titles.
-- "delete", "remove", "trash", "get rid of" a note (not a task) → use delete_note with query=<note name/keyword>. NEVER use this for tasks.
-- "delete", "remove", "trash", "get rid of" a task → use delete_task with query=<task name/keyword>. NEVER use this for notes.
-- "read", "open", "show me", "what does X say", "display", "pull up" a specific note → use get_note with query=<note name>.
-- "list my notes", "show notes", "recent notes", "browse notes", "notes tagged X" → use list_notes (with optional q or tags).
-- "tag X with Y", "add tag Y to X", "untag Y from X", "remove tag Y from X" → use update_note with tags=[Y] and tag_mode="add" or "remove".
-- search_web: user explicitly wants a quick factual answer from the web — current events, version
-  numbers, real-time facts ("what is the latest version of X", "who won the game last night").
-  ONLY use when the information is clearly not in their notes and they need something fresh from
-  the internet. Do NOT use for creative questions, brainstorming, game design, writing help, or
-  when the user is building on content they already have. Do NOT use when the user references
-  their own notes or prior research — use search_notes instead.
-- search_images: ONLY when the user explicitly asks to SEE, SHOW, DISPLAY, or VIEW an image or
-  photo. Trigger phrases: "show me a picture/photo/image of", "what does X look like",
-  "find a photo of", "display an image of". Do NOT use for general questions about appearance,
-  descriptions, or when the user just wants information without visual content.
-- research_topic: user wants a comprehensive, multi-section research note created from web sources.
-  Use whenever the user wants to deeply understand, learn about, or get a full written reference
-  on any subject — regardless of how they phrase it. The topic can be anything: technical subjects,
-  shopping decisions, comparisons, how things work, historical topics, etc.
-  The "topic" argument should capture the full subject matter of the request.
-  Prefer this over search_web when the user's request implies wanting thorough coverage rather than
-  a quick answer.
-- For creative/ideation requests ("think of", "come up with", "ideas for", "help me design",
-  "imagine", "brainstorm") use null (chat) — the main model answers these directly. Only route
-  to a tool if the user also explicitly asks to search, look something up, or create/save something.
-- If the user references notes or research the assistant previously created, prefer search_notes
-  (to retrieve the relevant note) or null (chat) so the main model can use its tools to find it.
-  Never use search_web when existing notes likely contain the answer.
-- "ack": one short, natural sentence confirming the action (tool path only). Vary phrasing — do not always start with "Let me". Omit (null) for chat-only responses.
-- Do NOT wrap the JSON in markdown code fences."""
-
-
-# Fast-path: "Research: <topic>" sent by the Research button.
-_RESEARCH_PREFIX = re.compile(r"^[Rr]esearch:\s+(.+)", re.DOTALL)
-
-# When the user refers to work the assistant previously did, they want the main
-# model to answer using existing context — not a web search.  Skip intent and
-# fall through to the streaming path so the model can call search_notes itself.
-_PRIOR_WORK_REFS = re.compile(
-    r"\b("
-    r"research (you|that you|we) did"
-    r"|research (you|that you|we) (made|created|compiled|wrote)"
-    r"|note (you|that you|we) (made|created|wrote)"
-    r"|notes (you|that you|we) (made|created|wrote)"
-    r"|using (your|the) research"
-    r"|based on (your|the) research"
-    r"|from (your|the) (research|note|notes)"
-    r")\b",
-    re.IGNORECASE,
-)
-
-
-async def classify_intent(
-    user_message: str,
-    tools: list[dict],
-    model: str,
-    history: list[dict] | None = None,
-) -> IntentResult:
-    """Classify user intent via a fast non-streaming LLM call.
-
-    history is a list of recent {role, content} messages (user/assistant only,
-    no system messages) for resolving anaphoric references.
-
-    Returns an IntentResult. On any failure, returns IntentResult(tool_name=None)
-    so the caller falls through to the normal streaming path.
-    """
-    if not tools:
-        return IntentResult()
-
-    # Fast-path: "Research: <topic>" is the canonical format sent by the Research
-    # button in the UI. It always means research_topic — skip the LLM call entirely.
-    valid_names = {t.get("function", {}).get("name") for t in tools}
-    if "research_topic" in valid_names:
-        m = _RESEARCH_PREFIX.match(user_message.strip())
-        if m:
-            topic = m.group(1).strip()
-            logger.info("Intent fast-path: 'Research:' prefix → research_topic, topic='%s'", topic[:80])
-            return IntentResult(
-                tool_name="research_topic",
-                arguments={"topic": topic},
-                confidence="high",
-                ack=f"I'll research that and compile a comprehensive note.",
-            )
-
-    # Fast-path: user references prior assistant work ("research you did", "note you
-    # made", etc.) — this is a request to use existing content, not search the web.
-    # Return no-tool so the main model answers conversationally with search_notes
-    # available if it needs to retrieve the note.
-    if _PRIOR_WORK_REFS.search(user_message):
-        logger.info(
-            "Intent fast-path: prior-work reference detected → skipping tool dispatch"
-        )
-        return IntentResult()
-
-    tool_summary = _build_tool_summary(tools)
-    today = date_type.today().isoformat()
-
-    messages: list[dict] = [
-        {
-            "role": "system",
-            "content": _SYSTEM_PROMPT_TEMPLATE.format(
-                today=today, tool_summary=tool_summary
-            ),
-        },
-    ]
-
-    # Inject recent history turns so the model can resolve references
-    if history:
-        for turn in history:
-            role = turn.get("role")
-            content = turn.get("content", "")
-            if role in ("user", "assistant") and content:
-                messages.append({"role": role, "content": content})
-
-    messages.append({"role": "user", "content": user_message})
-
-    try:
-        raw = await generate_completion(messages, model, max_tokens=350, num_ctx=4096)
-    except Exception:
-        logger.warning("Intent classification LLM call failed", exc_info=True)
-        return IntentResult()
-
-    return _parse_intent(raw, tools)
-
 
 def _parse_intent(raw: str, tools: list[dict]) -> IntentResult:
     """Parse the LLM's JSON response into an IntentResult."""
diff --git a/src/fabledassistant/services/research.py b/src/fabledassistant/services/research.py
index 03744b9..66bc7c0 100644
--- a/src/fabledassistant/services/research.py
+++ b/src/fabledassistant/services/research.py
@@ -25,7 +25,6 @@ async def run_research_pipeline(
     topic: str,
     user_id: int,
     model: str,
-    intent_model: str,
     buf=None,
 ) -> Note:
     """Full research pipeline: search → fetch → synthesize → create note.
@@ -36,7 +35,7 @@ async def run_research_pipeline(
     # Step 1: Generate sub-queries
     if buf is not None:
         buf.append_event("status", {"status": "Generating search queries..."})
-    queries = await _generate_sub_queries(topic, intent_model)
+    queries = await _generate_sub_queries(topic, model)
     logger.info("Research: generated %d sub-queries for topic '%s'", len(queries), topic)
 
     # Step 2: Search all queries in parallel (200 ms stagger to avoid hammering SearXNG)
@@ -118,8 +117,8 @@ async def run_research_pipeline(
     return note
 
 
-async def _generate_sub_queries(topic: str, intent_model: str) -> list[str]:
-    """Ask the intent model for focused search queries for the topic."""
+async def _generate_sub_queries(topic: str, model: str) -> list[str]:
+    """Ask the model for focused search queries for the topic."""
     messages = [
         {
             "role": "system",
@@ -135,7 +134,7 @@ async def _generate_sub_queries(topic: str, intent_model: str) -> list[str]:
         {"role": "user", "content": f"Topic: {topic}"},
     ]
     try:
-        raw = await generate_completion(messages, intent_model, max_tokens=200)
+        raw = await generate_completion(messages, model, max_tokens=200)
         raw = raw.strip()
         raw = re.sub(r"^```(?:json)?\s*", "", raw)
         raw = re.sub(r"\s*```$", "", raw)