Improve data indexing and prompt template for WDI

llm4data/llm/indicators/wdi_sql.py

@@ -34,6 +34,10 @@ def load_wdi_jsons(cls, wdi_jsons_dir: Path):
         Do bulk insert.
         """
 
+        # Drop and recreate the table
+        cls.__table__.drop(engine, checkfirst=True)
+        Base.metadata.create_all(engine)
+
         wdi_jsons_dir = Path(wdi_jsons_dir)
 
         if not wdi_jsons_dir.exists() or not wdi_jsons_dir.is_dir():
@@ -48,6 +52,10 @@ def load_wdi_jsons(cls, wdi_jsons_dir: Path):
             session.bulk_save_objects(wdi_objects)
             session.commit()
 
+        # # Create an index on the indicator column
+        # psql -U postgres -d wdi
+        # CREATE INDEX indicator_index ON wdi USING hash (indicator);
+
     @classmethod
     def run_sql(
         cls,

llm4data/prompts/indicators/wdi.py

@@ -19,7 +19,8 @@ def __init__(self, input_variables=None, template=None):
                 " Write an SQL query for the prompt: ```{{{{user_content}}}}```\n\n"
                 "table: {table}\n"
                 "fields: {fields}\n\n"
-                " Use the convention `:param` and not `?`."
+                "Only the indicator can parameterized and you must fill the rest."
+                " Use the convention `:indicator` and not `?`."
                 " Use country_iso3 when querying, use country in the result.\n\n"
                 "Use the last 10 years if no year is specified."
                 " Drop rows with no value.\n\n"

llm4data/scripts/indexing/create_field_index.py

@@ -0,0 +1,46 @@
+from llm4data.index import get_docs_index
+from llm4data import configs
+import fire
+
+
+def create_doc_index(field_name: str, field_schema: str):
+    """
+    Create a field index for the docs collection.
+
+    Args:
+        field_name (str): The name of the field to index.
+        field_schema (str): The schema of the field to index.
+
+    Returns:
+        dict: The response from the Qdrant server.
+
+    Examples:
+        >>> from llm4data.scripts.indexing.create_field_index import create_doc_index
+        >>> from llm4data import configs
+        >>> create_doc_index("document_description.title_statement.idno", "keyword")
+    """
+
+    # Check that the field name starts with "document_description." since that is the
+    # expected metadata key for the docs schema.
+    assert field_name.startswith("document_description."), f"Field name must start with 'document_description.' but got {field_name}"
+    doc_index = get_docs_index()
+
+    indexed = doc_index.client.create_payload_index(
+        doc_index.collection_name,
+        field_name=f"metadata.{configs.METADATA_KEY}.{field_name}",
+        field_schema=field_schema,
+    )
+
+    return indexed
+
+
+def main(data_type: str, field_name: str, field_schema: str):
+    if data_type == "docs":
+        create_doc_index(field_name, field_schema)
+    else:
+        raise ValueError(f"Unknown data type {data_type}")
+
+
+if __name__ == "__main__":
+    # python -m llm4data.scripts.indexing.create_field_index --data_type=docs --field_name="document_description.title_statement.idno" --field_schema=keyword
+    fire.Fire(main)

llm4data/scripts/indexing/indicators/load_indicators.py

@@ -57,6 +57,9 @@ def load_indicators(collection_dir: Path):
                 metadata=s.dict(exclude_none=True),
             )
 
+            with open(indexed_indicators_path, "a+") as f:
+                f.write(f"{indicator_path}\n")
+
         except KeyboardInterrupt:
             raise KeyboardInterrupt