Fix embedding size and correctly filter default models (#31)

* correctly filter default model * minor typing fix * fix chat mode streaming with non-multiturn columns * minor fix to ollama model, update to qwen2.5 3B * update changelog --------- Co-authored-by: deafnv <[email protected]>
EmbeddedLLM · Nov 29, 2024 · 5c3df39 · 5c3df39
1 parent b144455
commit 5c3df39
Show file tree

Hide file tree

Showing 36 changed files with 827 additions and 768 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -14,6 +14,25 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 
 - The version number mentioned here refers to the cloud version. For each release, all SDKs will have the same major and minor version, but their patch version may differ. For example, latest Python SDK might be `v0.2.0` whereas TS SDK might be `v0.2.1`, but both will be compatible with release `v0.2`.
 
+## [Unreleased]
+
+Backend - owl (API server)
+
+- Fix bge-small embedding size (1024 -> 384)
+- Correctly filter models at auth level
+- Fix ollama model deployment config
+
+Frontend
+
+- Added support for multiple multiturn columns in Chat table chat view.
+- Added multiturn chat toggle to column settings.
+
+Docker
+
+- Added Mac Apple Silicon `compose.mac.yml`
+- Update `ollama.yml` to use Qwen2.5 3B
+- Fix ollama default config
+
 ## [v0.3.1] (2024-11-26)
 
 This is a bug fix release for frontend code. SDKs are not affected.

diff --git a/clients/python/src/jamaibase/protocol.py b/clients/python/src/jamaibase/protocol.py
@@ -1719,46 +1719,27 @@ class ActionTableSchemaCreate(TableSchemaCreate):
 
 
 class AddActionColumnSchema(ActionTableSchemaCreate):
+    # TODO: Deprecate this
     pass
 
 
 class KnowledgeTableSchemaCreate(TableSchemaCreate):
+    # TODO: Maybe deprecate this and use EmbedGenConfig instead ?
     embedding_model: str
 
-    @model_validator(mode="after")
-    def check_cols(self) -> Self:
-        super().check_cols()
-        num_text_cols = sum(c.id.lower() in ("text", "title", "file id") for c in self.cols)
-        if num_text_cols != 0:
-            raise ValueError("Schema cannot contain column names: 'Text', 'Title', 'File ID'.")
-        return self
-
 
 class AddKnowledgeColumnSchema(TableSchemaCreate):
-    @model_validator(mode="after")
-    def check_cols(self) -> Self:
-        super().check_cols()
-        num_text_cols = sum(c.id.lower() in ("text", "title", "file id") for c in self.cols)
-        if num_text_cols != 0:
-            raise ValueError("Schema cannot contain column names: 'Text', 'Title', 'File ID'.")
-        return self
+    # TODO: Deprecate this
+    pass
 
 
 class ChatTableSchemaCreate(TableSchemaCreate):
-    @model_validator(mode="after")
-    def check_cols(self) -> Self:
-        super().check_cols()
-        num_text_cols = sum(c.id.lower() in ("user", "ai") for c in self.cols)
-        if num_text_cols != 2:
-            raise ValueError("Schema must contain column names: 'User' and 'AI'.")
-        return self
+    pass
 
 
 class AddChatColumnSchema(TableSchemaCreate):
-    @model_validator(mode="after")
-    def check_cols(self) -> Self:
-        super().check_cols()
-        return self
+    # TODO: Deprecate this
+    pass
 
 
 class TableMeta(TableBase):

diff --git a/clients/python/tests/oss/gen_table/test_table_ops.py b/clients/python/tests/oss/gen_table/test_table_ops.py
@@ -21,6 +21,7 @@
     "str": '"Arrival" is a 2016 science fiction film. "Arrival" è un film di fantascienza del 2016. 「Arrival」は2016年のSF映画です。',
 }
 KT_FIXED_COLUMN_IDS = ["Title", "Title Embed", "Text", "Text Embed", "File ID"]
+CT_FIXED_COLUMN_IDS = ["User"]
 
 TABLE_ID_A = "table_a"
 TABLE_ID_B = "table_b"
@@ -1430,6 +1431,21 @@ def test_kt_drop_invalid_columns(client_cls: Type[JamAI]):
                 )
 
 
+@flaky(max_runs=5, min_passes=1, rerun_filter=_rerun_on_fs_error_with_delay)
+@pytest.mark.parametrize("client_cls", CLIENT_CLS)
+def test_ct_drop_invalid_columns(client_cls: Type[JamAI]):
+    table_type = "chat"
+    jamai = client_cls()
+    with _create_table(jamai, table_type) as table:
+        assert isinstance(table, p.TableMetaResponse)
+        for col in CT_FIXED_COLUMN_IDS:
+            with pytest.raises(RuntimeError):
+                jamai.table.drop_columns(
+                    table_type,
+                    p.ColumnDropRequest(table_id=table.id, column_names=[col]),
+                )
+
+
 @flaky(max_runs=5, min_passes=1, rerun_filter=_rerun_on_fs_error_with_delay)
 @pytest.mark.parametrize("client_cls", CLIENT_CLS)
 @pytest.mark.parametrize("table_type", TABLE_TYPES)
@@ -1450,7 +1466,7 @@ def test_rename_columns(
         assert isinstance(table, p.TableMetaResponse)
         assert all(isinstance(c, p.ColumnSchema) for c in table.cols)
         # Test rename on empty table
-        table = jamai.rename_columns(
+        table = jamai.table.rename_columns(
             table_type,
             p.ColumnRenameRequest(table_id=table.id, column_map=dict(y="z")),
         )
@@ -1475,7 +1491,7 @@ def test_rename_columns(
         _add_row(jamai, table_type, False, data=dict(x="True", z="<dummy>"))
         # Test rename table with data
         # Test also auto gen config reference update
-        table = jamai.rename_columns(
+        table = jamai.table.rename_columns(
             table_type,
             p.ColumnRenameRequest(table_id=table.id, column_map=dict(x="a")),
         )
@@ -1503,14 +1519,14 @@ def test_rename_columns(
 
         # Repeated new column names
         with pytest.raises(RuntimeError):
-            jamai.rename_columns(
+            jamai.table.rename_columns(
                 table_type,
                 p.ColumnRenameRequest(table_id=table.id, column_map=dict(a="b", z="b")),
             )
 
         # Overlapping new and old column names
         with pytest.raises(RuntimeError):
-            jamai.rename_columns(
+            jamai.table.rename_columns(
                 table_type,
                 p.ColumnRenameRequest(table_id=table.id, column_map=dict(a="b", z="a")),
             )
@@ -1525,7 +1541,22 @@ def test_kt_rename_invalid_columns(client_cls: Type[JamAI]):
         assert isinstance(table, p.TableMetaResponse)
         for col in KT_FIXED_COLUMN_IDS:
             with pytest.raises(RuntimeError):
-                jamai.rename_columns(
+                jamai.table.rename_columns(
+                    table_type,
+                    p.ColumnRenameRequest(table_id=table.id, column_map={col: col}),
+                )
+
+
+@flaky(max_runs=5, min_passes=1, rerun_filter=_rerun_on_fs_error_with_delay)
+@pytest.mark.parametrize("client_cls", CLIENT_CLS)
+def test_ct_rename_invalid_columns(client_cls: Type[JamAI]):
+    table_type = "chat"
+    jamai = client_cls()
+    with _create_table(jamai, table_type) as table:
+        assert isinstance(table, p.TableMetaResponse)
+        for col in CT_FIXED_COLUMN_IDS:
+            with pytest.raises(RuntimeError):
+                jamai.table.rename_columns(
                     table_type,
                     p.ColumnRenameRequest(table_id=table.id, column_map={col: col}),
                 )
@@ -1582,7 +1613,7 @@ def test_reorder_columns(
         cols = [c.id for c in table.cols]
         assert cols == expected_order, cols
         # Test reorder empty table
-        table = jamai.reorder_columns(
+        table = jamai.table.reorder_columns(
             table_type,
             p.ColumnReorderRequest(table_id=TABLE_ID_A, column_names=column_names),
         )
@@ -1692,7 +1723,7 @@ def test_reorder_columns_invalid(
         else:
             raise ValueError(f"Invalid table type: {table_type}")
         with pytest.raises(RuntimeError, match="referenced an invalid source column"):
-            jamai.reorder_columns(
+            jamai.table.reorder_columns(
                 table_type,
                 p.ColumnReorderRequest(table_id=TABLE_ID_A, column_names=column_names),
             )

diff --git a/docker/compose.cpu.ollama.yml b/docker/compose.cpu.ollama.yml
@@ -29,12 +29,12 @@ services:
         echo 'ollama serve did not start in time'; \
         exit 1; \
         fi; \
-        ollama pull phi3.5 && ollama cp phi3.5 microsoft/Phi3.5-mini-instruct; \
+        ollama pull qwen2.5:3b && ollama cp qwen2.5:3b Qwen/Qwen2.5-3B-Instruct; \
         tail -f /dev/null",
       ]
     restart: unless-stopped
     healthcheck:
-      test: ["CMD", "sh", "-c", "ollama show microsoft/Phi3.5-mini-instruct || exit 1"]
+      test: ["CMD", "sh", "-c", "ollama show Qwen/Qwen2.5-3B-Instruct || exit 1"]
       interval: 20s
       timeout: 2s
       retries: 20

diff --git a/docker/ollama.yml b/docker/ollama.yml
@@ -1,4 +1,4 @@
 services:
   owl:
     environment:
-      - OWL_MODELS_CONFIG="models_ollama.json"
+      - OWL_MODELS_CONFIG=models_ollama.json
diff --git a/scripts/remove_cloud_modules.ps1 b/scripts/remove_cloud_modules.ps1
@@ -19,3 +19,4 @@ function quiet_rm($item)
 quiet_rm "services/app/ecosystem.config.cjs"
 quiet_rm "services/appecosystem.json"
 quiet_rm ".github/workflows/trigger-push-gh-image.yml"
+quiet_rm ".github/workflows/ci.cloud.yml"
diff --git a/scripts/remove_cloud_modules.sh b/scripts/remove_cloud_modules.sh
@@ -9,3 +9,4 @@ find . -type d -name "(cloud)" -exec rm -rf {} +
 rm -f services/app/ecosystem.config.cjs
 rm -f services/app/ecosystem.json
 rm -f .github/workflows/trigger-push-gh-image.yml
+rm -f .github/workflows/ci.cloud.yml
diff --git a/services/api/pyproject.toml b/services/api/pyproject.toml
@@ -132,7 +132,7 @@ dependencies = [
     "sqlmodel~=0.0.21",
     "srsly~=2.4.8",
     # starlette 0.38.3 and 0.38.4 seem to have issues with background tasks
-    "starlette==0.38.2",
+    "starlette~=0.41.3",
     "stripe~=9.12.0",
     "tantivy~=0.22.0",
     "tenacity~=8.5.0",

diff --git a/services/api/src/owl/configs/models_aipc.json b/services/api/src/owl/configs/models_aipc.json
@@ -132,7 +132,7 @@
       "id": "ellm/BAAI/bge-small-en-v1.5",
       "name": "ELLM BAAI BGE Small EN v1.5",
       "context_length": 512,
-      "embedding_size": 1024,
+      "embedding_size": 384,
       "languages": ["mul"],
       "capabilities": ["embed"],
       "deployments": [

diff --git a/services/api/src/owl/configs/models_ollama.json b/services/api/src/owl/configs/models_ollama.json
@@ -43,15 +43,15 @@
       ]
     },
     {
-      "id": "ellm/microsoft/Phi3.5-mini-instruct",
-      "name": "ELLM Phi3.5 mini instruct (3.8B)",
-      "context_length": 131072,
+      "id": "ellm/Qwen/Qwen2.5-3B-Instruct",
+      "name": "ELLM Qwen2.5 (3B)",
+      "context_length": 32000,
       "languages": ["en"],
       "capabilities": ["chat"],
       "deployments": [
         {
-          "litellm_id": "ollama_chat/microsoft/Phi3.5-mini-instruct",
-          "api_base": "http://ollama:11434",
+          "litellm_id": "openai/Qwen/Qwen2.5-3B-Instruct",
+          "api_base": "http://ollama:11434/v1",
           "provider": "ellm"
         }
       ]
@@ -62,7 +62,7 @@
       "id": "ellm/BAAI/bge-small-en-v1.5",
       "name": "ELLM BAAI BGE Small EN v1.5",
       "context_length": 512,
-      "embedding_size": 1024,
+      "embedding_size": 384,
       "languages": ["mul"],
       "capabilities": ["embed"],
       "deployments": [