Assistant improvements: Table annotations, and few-shot examples (#664)

* Table annotations
* Few-shot prompt examples
* View Assistant history
* Better 'relevant table' detection and UI
* Improved prompts
* Cmd+shift+F shortcut for formatting SQL

HISTORY.rst

@@ -7,13 +7,44 @@ This project adheres to `Semantic Versioning <https://semver.org/>`_.
 
 vNext
 ===========================
-* `#660`_: Userspace connection migration. This should be an invisible change, but represents a significant refactor of how connections function.
-Instead of a weird blend of DatabaseConnection models and underlying Django models (which were the original Explorer connections),
-this migrates all connections to DatabaseConnection models and implements proper foreign keys to them on the Query and QueryLog models.
-A data migration creates new DatabaseConnection models based on the configured settings.EXPLORER_CONNECTIONS.
-Going forward, admins can create new Django-backed DatabaseConnection models by registering the connection in EXPLORER_CONNECTIONS, and then creating a
-DatabaseConnection model using the Django admin or the user-facing /connections/new/ form, and entering the Django DB alias and setting the connection type to "Django Connection"
-
+* `#664`_: Improvements to the AI SQL Assistant:
+
+  - Table Annotations: Write persistent table annotations with descriptive information that will get injected into the
+    prompt for the assistant. For example, if a table is commonly joined to another table through a non-obvious foreign
+    key, you can tell the assistant about it in plain english, as an annotation to that table. Every time that table is
+    deemed 'relevant' to an assistant request, that annotation will be included alongside the schema and sample data.
+  - Few-Shot Examples: Using the small checkbox on the bottom-right of any saved queries, you can designate certain
+    queries as 'few shot examples". When making an assistant request, any designated few-shot examples that reference
+    the same tables as your assistant request will get included as 'reference sql' in the prompt for the LLM.
+  - Autocomplete / multiselect when selecting tables info to send to the SQL Assistant. Much easier and more keyboard
+    focused.
+  - Relevant tables are added client-side visually, in real time, based on what's in the SQL editor and/or any tables
+    mentioned in the assistant request. The dependency on sql_metadata is therefore removed, as server-side SQL parsing
+    is no longer necessary.
+  - Ability to view Assistant request/response history.
+  - Improved system prompt that emphasizes the particular SQL dialect being used.
+  - Addresses issue #657.
+
+* `#660`_: Userspace connection migration.
+
+  - This should be an invisible change, but represents a significant refactor of how connections function. Instead of a
+    weird blend of DatabaseConnection models and underlying Django models (which were the original Explorer
+    connections), this migrates all connections to DatabaseConnection models and implements proper foreign keys to them
+    on the Query and QueryLog models. A data migration creates new DatabaseConnection models based on the configured
+    settings.EXPLORER_CONNECTIONS. Going forward, admins can create new Django-backed DatabaseConnection models by
+    registering the connection in EXPLORER_CONNECTIONS, and then creating a DatabaseConnection model using the Django
+    admin or the user-facing /connections/new/ form, and entering the Django DB alias and setting the connection type
+    to "Django Connection".
+  - The Query.connection and QueryLog.connection fields are deprecated and will be removed in a future release. They
+    are kept around in this release in case there is an unforeseen issue with the migration. Preserving the fields for
+    now ensures there is no data loss in the event that a rollback to an earlier version is required.
+
+* Fixed a bug when validating connections to uploaded files. Also added basic locking when downloading files from S3.
+
+* Keyboard shortcut for formatting the SQL in the editor.
+
+  - Cmd+Shift+F (Windows: Ctrl+Shift+F)
+  - The format button has been moved tobe a small icon towards the bottom-right of the SQL editor.
 
 `5.2.0`_ (2024-08-19)
 ===========================
@@ -643,6 +674,8 @@ Initial Release
 .. _#651: https://github.com/explorerhq/sql-explorer/pull/651
 .. _#659: https://github.com/explorerhq/sql-explorer/pull/659
 .. _#662: https://github.com/explorerhq/sql-explorer/pull/662
+.. _#660: https://github.com/explorerhq/sql-explorer/pull/660
+.. _#664: https://github.com/explorerhq/sql-explorer/pull/664
 
 .. _#269: https://github.com/explorerhq/sql-explorer/issues/269
 .. _#288: https://github.com/explorerhq/sql-explorer/issues/288

docs/features.rst

@@ -5,7 +5,19 @@ SQL Assistant
 -------------
 - Built in integration with OpenAI (or the LLM of your choosing)
   to quickly get help with your query, with relevant schema
-  automatically injected into the prompt. Simple, effective.
+  automatically injected into the prompt.
+- The assistant tries hard to get relevant context into the prompt to the LLM, alongside your explicit request. You
+  can choose tables to include explicitly (and any tables you are reference in your SQL you will see get included as
+  well). When a table is "included", the prompt will include the schema of the table, 3 sample rows, any Table
+  Annotations you have added, and any designated "few shot examples". More on each of those below.
+- Table Annotations: Write persistent table annotations with descriptive information that will get injected into the
+  prompt for the assistant. For example, if a table is commonly joined to another table through a non-obvious foreign
+  key, you can tell the assistant about it in plain english, as an annotation to that table. Every time that table is
+  deemed 'relevant' to an assistant request, that annotation will be included alongside the schema and sample data.
+- Few-shot examples: Using the small checkbox on the bottom-right of any saved query, you can designate queries as
+  "Assistant Examples". When making an assistant request, the 'included tables' are intersected with tables referenced
+  by designated Example queries, and those queries are injected into the prompt, and the LLM is told that that these
+  are good reference queries.
 
 Database Support
 ----------------
@@ -222,8 +234,7 @@ Power tips
   view.
 - Command+Enter and Ctrl+Enter will execute a query when typing in
   the SQL editor area.
-- Hit the "Format" button to format and clean up your SQL (this is
-  non-validating -- just formatting).
+- Cmd+Shift+F (Windows: Ctrl+Shift+F) to format the SQL in the editor.
 - Use the Query Logs feature to share one-time queries that aren't
   worth creating a persistent query for. Just run your SQL in the
   playground, then navigate to ``/logs`` and share the link

explorer/admin.py

@@ -7,7 +7,7 @@
 
 @admin.register(Query)
 class QueryAdmin(admin.ModelAdmin):
-    list_display = ("title", "description", "created_by_user",)
+    list_display = ("title", "description", "created_by_user", "few_shot")
     list_filter = ("title",)
     raw_id_fields = ("created_by_user",)
     actions = [generate_report_action()]

explorer/app_settings.py

@@ -152,11 +152,17 @@
 EXPLORER_AI_API_KEY = getattr(settings, "EXPLORER_AI_API_KEY", None)
 
 EXPLORER_ASSISTANT_BASE_URL = getattr(settings, "EXPLORER_ASSISTANT_BASE_URL", "https://api.openai.com/v1")
+
+# Deprecated. Will be removed in a future release. Please use EXPLORER_ASSISTANT_MODEL_NAME instead
 EXPLORER_ASSISTANT_MODEL = getattr(settings, "EXPLORER_ASSISTANT_MODEL",
                                    # Return the model name and max_tokens it supports
                                    {"name": "gpt-4o",
                                     "max_tokens": 128000})
 
+EXPLORER_ASSISTANT_MODEL_NAME = getattr(settings, "EXPLORER_ASSISTANT_MODEL_NAME",
+                                    EXPLORER_ASSISTANT_MODEL["name"])
+
+
 EXPLORER_DB_CONNECTIONS_ENABLED = getattr(settings, "EXPLORER_DB_CONNECTIONS_ENABLED", False)
 EXPLORER_USER_UPLOADS_ENABLED = getattr(settings, "EXPLORER_USER_UPLOADS_ENABLED", False)
 EXPLORER_PRUNE_LOCAL_UPLOAD_COPY_DAYS_INACTIVITY = getattr(settings,

explorer/assistant/forms.py

@@ -0,0 +1,42 @@
+from django import forms
+from explorer.assistant.models import TableDescription
+from explorer.ee.db_connections.utils import default_db_connection
+
+
+class TableDescriptionForm(forms.ModelForm):
+    class Meta:
+        model = TableDescription
+        fields = "__all__"
+        widgets = {
+            "database_connection": forms.Select(attrs={"class": "form-select"}),
+            "description": forms.Textarea(attrs={"class": "form-control", "rows": 3}),
+        }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if not self.instance.pk:  # Check if this is a new instance
+            # Set the default value for database_connection
+            self.fields["database_connection"].initial = default_db_connection()
+
+        if self.instance and self.instance.table_name:
+            choices = [(self.instance.table_name, self.instance.table_name)]
+        else:
+            choices = []
+
+        f = forms.ChoiceField(
+            choices=choices,
+            widget=forms.Select(attrs={"class": "form-select", "data-placeholder": "Select table"})
+        )
+
+        # We don't actually care about validating the 'choices' that the ChoiceField does by default.
+        # Really we are just using that field type in order to get a valid pre-populated Select widget on the client
+        # But also it can't be blank!
+        def valid_value_new(v):
+            return bool(v)
+
+        f.valid_value = valid_value_new
+
+        self.fields["table_name"] = f
+
+        if self.instance and self.instance.table_name:
+            self.fields["table_name"].initial = self.instance.table_name

explorer/assistant/models.py

@@ -1,5 +1,6 @@
 from django.db import models
 from django.conf import settings
+from explorer.ee.db_connections.models import DatabaseConnection
 
 
 class PromptLog(models.Model):
@@ -8,6 +9,7 @@ class Meta:
         app_label = "explorer"
 
     prompt = models.TextField(blank=True)
+    user_request = models.TextField(blank=True)
     response = models.TextField(blank=True)
     run_by_user = models.ForeignKey(
         settings.AUTH_USER_MODEL,
@@ -19,3 +21,18 @@ class Meta:
     duration = models.FloatField(blank=True, null=True)  # seconds
     model = models.CharField(blank=True, max_length=128, default="")
     error = models.TextField(blank=True, null=True)
+    database_connection = models.ForeignKey(to=DatabaseConnection, on_delete=models.SET_NULL, blank=True, null=True)
+
+
+class TableDescription(models.Model):
+
+    class Meta:
+        app_label = "explorer"
+        unique_together = ("database_connection", "table_name")
+
+    database_connection = models.ForeignKey(to=DatabaseConnection, on_delete=models.CASCADE)
+    table_name = models.CharField(max_length=512)
+    description = models.TextField()
+
+    def __str__(self):
+        return f"{self.database_connection.alias} - {self.table_name}"

explorer/assistant/urls.py

@@ -0,0 +1,16 @@
+from django.urls import path
+from explorer.assistant.views import (TableDescriptionListView,
+                                      TableDescriptionCreateView,
+                                      TableDescriptionUpdateView,
+                                      TableDescriptionDeleteView,
+                                      AssistantHelpView,
+                                      AssistantHistoryApiView)
+
+assistant_urls = [
+    path("assistant/", AssistantHelpView.as_view(), name="assistant"),
+    path("assistant/history/", AssistantHistoryApiView.as_view(), name="assistant_history"),
+    path("table-descriptions/", TableDescriptionListView.as_view(), name="table_description_list"),
+    path("table-descriptions/new/", TableDescriptionCreateView.as_view(), name="table_description_create"),
+    path("table-descriptions/<int:pk>/update/", TableDescriptionUpdateView.as_view(), name="table_description_update"),
+    path("table-descriptions/<int:pk>/delete/", TableDescriptionDeleteView.as_view(), name="table_description_delete"),
+]