Merge pull request #81 from Tanzania-AI-Community/feature/tool-calling

Feature/tool calling

Author: jurmy24

.pre-commit-config.yaml

@@ -1,27 +1,27 @@
-# See https://pre-commit.com for more information
-# See https://pre-commit.com/hooks.html for more hooks
-fail_fast: true
-default_stages: [pre-commit]
-
-repos:
-  - repo: https://github.com/psf/black
-    rev: 24.10.0
-    hooks:
-      - id: black
-        args: [--config, pyproject.toml]
-        types: [python]
-
-  - repo: https://github.com/charliermarsh/ruff-pre-commit
-    rev: "v0.7.3"
-    hooks:
-      - id: ruff
-        args: [--fix, --exit-non-zero-on-fix]
-
-  - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
-    hooks:
-      - id: check-toml
-      - id: check-yaml
-      - id: detect-private-key
-      - id: end-of-file-fixer
-      - id: trailing-whitespace
+# See https://pre-commit.com for more information
+# See https://pre-commit.com/hooks.html for more hooks
+fail_fast: true
+default_stages: [pre-commit]
+
+repos:
+  - repo: https://github.com/psf/black
+    rev: 24.10.0
+    hooks:
+      - id: black
+        args: [--config, pyproject.toml]
+        types: [python]
+
+  - repo: https://github.com/charliermarsh/ruff-pre-commit
+    rev: "v0.7.3"
+    hooks:
+      - id: ruff
+        args: [--fix, --exit-non-zero-on-fix]
+
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v5.0.0
+    hooks:
+      - id: check-toml
+      - id: check-yaml
+      - id: detect-private-key
+      - id: end-of-file-fixer
+      - id: trailing-whitespace

app/assets/prompts/exercise_generator_system

@@ -1,13 +1,8 @@
-You are a skilled Tanzanian secondary school teacher that generates questions or exercises for Tanzanian Form 2 geography students based on the request made by the user. Use the provided context from the textbook to ensure that the questions you generate are grounded in the course content. Given the context information and not prior knowledge, follow the query instructions provided by the user. Don't generate questions if the query topic from the user is not related to the course content. Begin your response immediately with the question.
+You are a skilled Tanzanian secondary school teacher that generates questions or exercises for students based on the request made by the user. This exercise is for {class_info} students. Use the provided context from the textbook to ensure that the questions you generate are grounded in the course content. Take inspiration from the example exercises from the textbook if they are provided to you. Given the context information and not prior knowledge, follow the query instructions provided by the user. Don't generate questions if the query topic from the user is not related to the course content. Begin your response immediately with the question.
 
 EXAMPLE INTERACTION:
 user: Follow these instructions (give me short answer question on Tanzania's mining industry)
 Context information is below.
-
-START
-
 Tanzania has many minerals that it trades with to other countries...etc.
 
-END
-
 assistant: List three minerals that Tanzania exports.

app/assets/prompts/twiga_system

@@ -1,12 +1,17 @@
-You are Twiga, a WhatsApp bot developed by the Tanzania AI Community specifically for secondary school teachers in Tanzania. Your role is to support teachers by providing accurate, curriculum-aligned educational assistance in a professional and supportive tone.
+You are Twiga, a WhatsApp bot developed by the Tanzania AI Community for secondary school teachers in Tanzania. The curriculum is the Tanzanian National Curriculum, developed by the Tanzanian Institute of Education (TIE). The students are assessed in NECTA examinations, which cover the curriculum. TIE are also the writers of the textbooks you use. Your role is to support the teachers by providing accurate, curriculum-aligned educational assistance. You are talking to {user_name} who teaches {class_info}.
 
-You are talking to {user_name} who teaches {class_info}
+Follow these instructions:
 
-Note that os2 refers to Ordinary Secondary Level 2, which is equivalent to Form 2 in the Tanzanian education system.
+- You only communicate in english
+- Use the tools you have available
+- Be clear and concise, since your messages are communicated on WhatsApp
+- Ask the teacher for additional information or clarification if its needed
+- Do not generate educational content if they are not provided by your tools
+- If the tool has an error or does not fulfill the user request, tell the user
+- Only help the teacher with subject related matter
 
-Follow these core guidelines:
+Here are your capabilities:
 
-Guidelines:
-Use Available Tools: For subject-related queries, refer to the tools available to you, unless the query is straightforward or involves general knowledge.
-Clarity and Conciseness: Ensure all responses are clear, concise, and easy to understand.
-Seek Clarification: If a query is unclear, kindly ask the user for additional details to provide a more accurate response.
+1. Searching the textbooks to answer course-related questions
+2. Generating example exercises or questions based on a specific course-related topic
+3. General tips and support

app/assets/strings/english.yml

@@ -20,8 +20,8 @@ settings:
   class_subject_info: "Classes and Subjects"
 
 tools:
-  exercise_generator: "🏋️ Generating an exercise using the course literature, please hold..."
-  search_knowledge: "🔍 Searching for relevant knowledge, please hold..."
+  generate_exercise: "📑 Generating exercises from the course content, please hold..."
+  search_knowledge: "📚 Searching the course content, please hold..."
 
 flows:
   start_onboarding_header: "Start onboarding to Twiga 🦒"

app/database/db.py

@@ -26,11 +26,24 @@ async def get_or_create_user(wa_id: str, name: Optional[str] = None) -> User:
     """
     Get existing user or create new one if they don't exist.
     Handles all database operations and error logging.
+    This uses a lot of eager loading to get the class information from the user.
     """
+    # TODO: Remove eager loading if too slow
     async with get_session() as session:
         try:
             # First try to get existing user
             statement = select(User).where(User.wa_id == wa_id).with_for_update()
+            # First try to get existing user
+            statement = (
+                select(User)
+                .where(User.wa_id == wa_id)
+                .options(
+                    selectinload(User.taught_classes)  # type: ignore
+                    .selectinload(TeacherClass.class_)  # type: ignore
+                    .selectinload(Class.subject_)  # type: ignore
+                )
+                .with_for_update()
+            )
             result = await session.execute(statement)
             user = result.scalar_one_or_none()
             if user:
@@ -181,6 +194,37 @@ async def vector_search(query: str, n_results: int, where: dict) -> List[Chunk]:
             raise Exception(f"Failed to search for knowledge: {str(e)}")
 
 
+async def get_class_resources(class_id: int) -> Optional[List[int]]:
+    """
+    Get all resource IDs accessible to a class.
+    Uses a single optimized SQL query with proper indexing.
+    """
+    async with get_session() as session:
+        try:
+            # Use text() for a more efficient raw SQL query
+            query = text(
+                """
+                SELECT DISTINCT resource_id
+                FROM classes_resources
+                WHERE class_id = :class_id
+                """
+            )
+
+            result = await session.execute(query, {"class_id": class_id})
+            resource_ids = [row[0] for row in result.fetchall()]
+
+            if not resource_ids:
+                logger.warning(f"No resources found for class {class_id}")
+                return None
+
+            logger.debug(f"Found resources {resource_ids} for class {class_id}")
+            return resource_ids
+
+        except Exception as e:
+            logger.error(f"Failed to get resources for class {class_id}: {str(e)}")
+            raise Exception(f"Failed to get class resources: {str(e)}")
+
+
 async def get_user_resources(user: User) -> Optional[List[int]]:
     """
     Get all resource IDs accessible to a user through their class assignments.

app/database/models.py

@@ -114,6 +114,22 @@ def formatted_class_info(self) -> str:
 
         return ClassInfo(classes=self.class_info).format_readable()
 
+    @property
+    def class_name_to_id_map(self) -> Dict[str, int]:
+        """
+        Returns a mapping of class names (with subject and form) to their IDs.
+        Example: {"Geography Form 2": 123, "Geography Form 3": 456}
+        """
+
+        if not self.taught_classes:
+            return {}
+
+        # TODO: This can be done in a better way without suppressing the type
+        return {
+            f"{class_.class_.subject_.name.capitalize()} {enums.GradeLevel(class_.class_.grade_level).display_format}": class_.class_.id
+            for class_ in self.taught_classes
+        }  # type: ignore
+
 
 class Subject(SQLModel, table=True):
     __tablename__ = "subjects"  # type: ignore
@@ -205,7 +221,7 @@ class Message(SQLModel, table=True):
 
     def to_api_format(self) -> Dict[str, Any]:
         """Convert message to OpenAI API format"""
-        message: Dict[str, Any] = {"role": self.role}
+        message: Dict[str, Any] = {"role": self.role.value}
         if self.tool_calls and len(self.tool_calls) > 0:
             message["tool_calls"] = self.tool_calls
             message["content"] = None