update

.gitignore

@@ -160,3 +160,10 @@ cython_debug/
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 #.idea/
+.DS_Store
+README.p.md
+ai-search-demo/.DS_Store
+ai-search-demo/storage/
+ai-search-demo/example_data/
+ml-stages/data/
+ml-stages/script/

ai-search-demo/README.en.md

@@ -63,6 +63,11 @@ python ai_search_demo/evaluate_synthetic_data.py create-synthetic-dataset ./exam
 python ai_search_demo/evaluate_synthetic_data.py evaluate-on-synthetic-dataset koml/smart-hr-synthetic-data-single-image-multiple-queries --collection-name smart-hr-synthetic-data-single-image-multiple-queries
 ```
 
+## Demo data
+
+- [SmartHR](https://smarthr.jp/know-how/ebook/tv-campaign/)
+- VC Reports: [InfraRed](https://www.redpoint.com/infrared/report/) & [024: The State of Generative AI in the Enterprise](https://menlovc.com/2024-the-state-of-generative-ai-in-the-enterprise/)
+
 ## Architecture 
 
 High-level diagram of the system.
@@ -135,6 +140,14 @@ sequenceDiagram
 
 ## LLM inference 
 
+
+Setup
+
+```
+pip install modal
+modal setup
+```
+
 Download models
 
 ```
@@ -147,6 +160,6 @@ modal run llm-inference/llm_serving_load_models.py --model-name vidore/colqwen2-
 Deploy models
 
 ```
-modal deploy llm-inference/llm_serving.py 
+modal deploy llm-inference/llm_serving.py
 modal deploy llm-inference/llm_serving_colpali.py
 ```

ai-search-demo/README.md

@@ -134,6 +134,14 @@ sequenceDiagram
 
 ## LLM推論
 
+セットアップ
+
+```
+pip install modal
+modal setup
+```
+
+
 モデルのダウンロード
 
 ```
@@ -146,6 +154,6 @@ modal run llm-inference/llm_serving_load_models.py --model-name vidore/colqwen2-
 モデルをデプロイする
 
 ```
-modal deploy llm-inference/llm_serving.py 
+modal deploy llm-inference/llm_serving.py
 modal deploy llm-inference/llm_serving_colpali.py
 ```

ai-search-demo/ai_search_demo/ui.py

@@ -107,7 +107,7 @@ def ai_search():
 
                     # Display VLLM output in the second column
                     with col2:
-                        st.markdown("<h3 style='color:green;'>Relevant Images</h3>", unsafe_allow_html=True)
+                        st.markdown("<h3 style='color:green;'>Interpretation with LLM</h3>", unsafe_allow_html=True)
                         for image_data, _, _, _ in search_results_data:
                             with st.spinner("Processing with VLLM..."):
                                 vllm_output = call_vllm(image_data)

ai-search-demo/llm-inference/llm_serving.py

@@ -28,6 +28,7 @@
 @app.function(
     image=vllm_image,
     gpu=modal.gpu.H100(count=N_GPU),
+    keep_warm=1,
     container_idle_timeout=5 * MINUTES,
     timeout=24 * HOURS,
     allow_concurrent_inputs=1000,

ai-search-demo/llm-inference/llm_serving_colpali.py

@@ -27,6 +27,7 @@
 @app.function(
     image=vllm_image,
     gpu=modal.gpu.H100(count=N_GPU),
+    keep_warm=1,
     container_idle_timeout=5 * MINUTES,
     timeout=24 * HOURS,
     allow_concurrent_inputs=1000,

ai-search-demo/llm-inference/llm_serving_load_models.py

@@ -5,8 +5,6 @@
 DEFAULT_NAME = "Qwen/Qwen2.5-7B-Instruct"
 DEFAULT_REVISION = "bb46c15ee4bb56c5b63245ef50fd7637234d6f75"
 
-# Qwen/Qwen2-VL-7B-Instruct
-
 
 volume = modal.Volume.from_name("models", create_if_missing=True)
 

ml-stages/Dockerfile

@@ -0,0 +1,14 @@
+FROM pytorch/pytorch:2.5.1-cuda12.1-cudnn9-devel
+ARG DEBIAN_FRONTEND=noninteractive
+
+ENV TZ=America/Los_Angeles
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential \
+    ffmpeg \
+    git \
+    git-lfs
+
+RUN pip install --upgrade pip
+COPY requirements.txt requirements.txt
+RUN pip install -r requirements.txt

ml-stages/README.en.md

@@ -0,0 +1,4 @@
+# ML  Stages
+
+![alt text](./docs/end2end.jpg)
+Based on this design document: [SmartHR LLM Workshop, December 2nd, 2024](https://docs.google.com/document/d/1xvJiWK8nvtC8Ek34X5Uk1BCCa_GvYqej-8tjk8o8obE/edit?tab=t.0)

ml-stages/README.md

@@ -0,0 +1,4 @@
+# 機械学習のステージ
+
+![alt text](./docs/end2end.jpg)
+この設計文書に基づいています: [SmartHR LLM ワークショップ、2024年12月2日](https://docs.google.com/document/d/1xvJiWK8nvtC8Ek34X5Uk1BCCa_GvYqej-8tjk8o8obE/edit?tab=t.0)

ml-stages/app.py

@@ -0,0 +1,98 @@
+import streamlit as st
+import duckdb
+from openai import OpenAI
+
+
+from pydantic import BaseModel
+from traceloop.sdk import Traceloop
+from traceloop.sdk.decorators import workflow
+
+Traceloop.init(app_name="duckdb-project", disable_batch=True)
+
+class Text2SQLSample(BaseModel):
+    query: str
+
+# @observe
+@workflow(name="text2sql")
+def text2sql(user_prompt: str, schema: str, table_name: str) -> str:
+    model: str = "gpt-4o"
+    client = OpenAI()
+
+    prompt = f"""
+    Write the corresponding SQL query based on user prompt and database schema:
+
+    - user prompt: {user_prompt}
+    - database schema: {schema}
+    Return only JSON.
+
+    Table name is {table_name}
+    """
+
+    chat_completion = client.beta.chat.completions.parse(
+    messages=[
+        {
+            "role": "system",
+            "content": "You are DuckDB and SQL expert.",
+        },
+        {
+            "role": "user",
+            "content": prompt,
+        },
+    ],
+    model=model,
+    response_format=Text2SQLSample,
+    temperature=1,
+    )
+    query = chat_completion.choices[0].message.parsed.query
+    return query
+
+# Initialize DuckDB connection
+conn = duckdb.connect()
+
+# Load the httpfs extension for hf:// paths
+conn.execute("INSTALL httpfs;")
+conn.execute("LOAD httpfs;")
+
+# Title
+st.title("Hugging Face Dataset Query with DuckDB")
+
+# Input for Hugging Face dataset link
+hf_link = st.text_input(
+    "Enter the Hugging Face dataset link (hf://...):",
+    "hf://datasets/UCSC-VLAA/Recap-DataComp-1B/data/train_data/train-00000-of-02719.parquet",
+)
+
+# Input for NLP query
+nlp_query = st.text_area("Enter your query in natural language:", "Show me the first 10 rows.")
+
+# Execute query when button is clicked
+if st.button("Run Query"):
+
+    try:
+        # Use DESCRIBE to get schema
+        schema_query = f"DESCRIBE SELECT * FROM '{hf_link}'"
+        schema_df = conn.execute(schema_query).df()
+        st.write("### Dataset Schema:")
+        st.dataframe(schema_df)
+
+        # Convert schema_df to string format suitable for text2sql
+        schema_str = "\n".join([f"{row['column_name']} ({row['column_type']})" for index, row in schema_df.iterrows()])
+
+    except Exception as e:
+        st.error(f"An error occurred while fetching schema: {e}")
+        schema_str = ""  # Ensure schema_str is defined
+
+    try:
+        # Convert NLP query to SQL using text2sql
+        sql_query = text2sql(nlp_query, schema_str, table_name=hf_link)
+        st.write("### Generated SQL Query:")
+        st.code(sql_query, language='sql')
+
+        # sql_query = sql_query.replace("table_name", hf_link)
+        # Execute the SQL query
+        df = conn.execute(sql_query).df()
+        # Display the results
+        st.write("### Query Results:")
+        st.dataframe(df)
+    except Exception as e:
+        st.error(f"An error occurred: {e}")

ml-stages/data.py

@@ -0,0 +1,140 @@
+from typing import Dict, List
+
+import argilla as rg
+import typer
+from datasets import Dataset, load_dataset
+from openai import OpenAI
+from pydantic import BaseModel
+from retry import retry
+from tqdm import tqdm
+
+
+class Text2SQLSample(BaseModel):
+    prompt: str
+    schema: str
+    query: str
+
+client = rg.Argilla(api_url="http://0.0.0.0:6900", api_key="argilla.apikey")
+# argilla / 12345678 / argilla.apikey
+
+def create_text2sql_dataset(dataset_name: str, data_samples: Dataset | List[Dict]):
+    guidelines = """
+    Please examine the given DuckDB SQL question and context. 
+    Write the correct DuckDB SQL query that accurately answers the question based on the context provided. 
+    Ensure the query follows DuckDB SQL syntax and logic correctly.
+    """
+
+    settings = rg.Settings(
+        guidelines=guidelines,
+        fields=[
+            rg.TextField(
+                name="prompt",
+                title="Prompt",
+                use_markdown=False,
+            ),
+            rg.TextField(
+                name="schema",
+                title="Schema",
+                use_markdown=True,
+            ),
+            rg.TextField(
+                name="query",
+                title="Query",
+                use_markdown=True,
+            ),            
+        ],
+        questions=[
+            rg.TextQuestion(
+                name="sql",
+                title="Please write SQL for this query",
+                description="Please write SQL for this query",
+                required=True,
+                use_markdown=True,
+            )
+        ],
+    )
+
+    if 'admin' not in [x.name for x in client.workspaces.list()]:
+        workspace = rg.Workspace(name="admin")
+        workspace.create()
+
+    dataset = rg.Dataset(
+        name=dataset_name,
+        workspace="admin",
+        settings=settings,
+        client=client,
+    )
+    dataset.create()
+    records = []
+    for idx in range(len(data_samples)):
+        x = rg.Record(
+            fields={
+                "prompt": data_samples[idx]["prompt"],
+                "schema": data_samples[idx]["schema"],
+                "query": data_samples[idx]["query"],
+            },
+        )
+        records.append(x)
+    dataset.records.log(records, batch_size=1000)
+
+def upload_duckdb_text2sql():
+    dataset_name = "motherduckdb/duckdb-text2sql-25k"
+    raw_dataset = load_dataset(dataset_name, split="train")
+    raw_datasets = raw_dataset.train_test_split(test_size=0.05, seed=42)
+
+    raw_dataset.to_json(path_or_buf='./data/duckdb-text2sql-25k.json')
+    raw_datasets['train'].to_json(path_or_buf='./data/train.json')
+    raw_datasets['test'].to_json(path_or_buf='./data/test.json')
+
+    create_text2sql_dataset(dataset_name='duckdb-text2sql-train', data_samples=raw_datasets['train'].to_list())
+    create_text2sql_dataset(dataset_name='duckdb-text2sql-test', data_samples=raw_datasets['test'].to_list())
+
+
+@retry(tries=3, delay=1)
+def generate_synthetic_example() -> Dict[str, str]:
+    client = OpenAI()
+
+    prompt = """
+    Generate a example for text2sql task for DuckDB database: 
+    The example should include 
+    - schema: a valid database schema
+    - prompt: a typical user question related to this table prompt
+    - query: the corresponding SQL query to answer user prompt.
+    Return only JSON. Use Japanese language for prompt (user question).
+    """
+
+    chat_completion = client.beta.chat.completions.parse(
+        messages=[
+            {
+                "role": "system",
+                "content": "You are DuckDB and SQL expert.",
+            },
+            {
+                "role": "user",
+                "content": prompt,
+            },
+        ],
+        model="gpt-4o",
+        response_format=Text2SQLSample,
+        temperature=1,
+    )
+    sample = chat_completion.choices[0].message.parsed
+    return sample.model_dump()
+
+
+def create_text2sql_dataset_synthetic(num_samples: int = 10):
+
+
+    samples = []
+    for _ in tqdm(range(num_samples)):
+        sample = generate_synthetic_example()
+        samples.append(sample)
+
+    dataset_name = "duckdb-text2sql-synthetic-jp"
+    create_text2sql_dataset(dataset_name=dataset_name, data_samples=samples)
+
+if __name__ == "__main__":
+    app = typer.Typer()
+    app.command()(upload_duckdb_text2sql)
+    app.command()(create_text2sql_dataset_synthetic)
+    app()