Skip to content

armgit / txtai

Public Imported from GitHub
0

💡 All-in-one AI framework for semantic search, LLM orchestration and language model workflows

Add template task, closes #448

davidmezzetticb9e5b4acf71
diff --git a/src/python/txtai/workflow/base.py b/src/python/txtai/workflow/base.py
index 45e39c59..39924881 100644
--- a/src/python/txtai/workflow/base.py
+++ b/src/python/txtai/workflow/base.py
@@ -20,7 +20,6 @@ from .execute import Execute
# Logging configuration
logger = logging.getLogger(__name__)
-logger.setLevel(logging.INFO)
class Workflow:
@@ -162,7 +161,8 @@ class Workflow:
"""
# Run elements through each task
- for task in self.tasks:
+ for x, task in enumerate(self.tasks):
+ logger.debug("Running Task #%d", x)
elements = task(elements, executor)
# Yield results processed by all tasks
diff --git a/src/python/txtai/workflow/task/__init__.py b/src/python/txtai/workflow/task/__init__.py
index 40a913e0..80f7fad5 100644
--- a/src/python/txtai/workflow/task/__init__.py
+++ b/src/python/txtai/workflow/task/__init__.py
@@ -11,5 +11,6 @@ from .image import ImageTask
from .retrieve import RetrieveTask
from .service import ServiceTask
from .storage import StorageTask
+from .template import ExtractorTask, TemplateTask
from .url import UrlTask
from .workflow import WorkflowTask
diff --git a/src/python/txtai/workflow/task/base.py b/src/python/txtai/workflow/task/base.py
index 8558c0f6..5ef41ef7 100644
--- a/src/python/txtai/workflow/task/base.py
+++ b/src/python/txtai/workflow/task/base.py
@@ -2,12 +2,16 @@
Task module
"""
+import logging
import re
import types
import numpy as np
import torch
+# Logging configuration
+logger = logging.getLogger(__name__)
+
class Task:
"""
@@ -207,7 +211,7 @@ class Task:
"""
# Extract data from (id, data, tag) formatted elements
- if (self.unpack or force) and isinstance(element, tuple):
+ if (self.unpack or force) and isinstance(element, tuple) and len(element) > 1:
return element[1]
return element
@@ -225,7 +229,7 @@ class Task:
"""
# Pack data into (id, data, tag) formatted elements
- if self.unpack and isinstance(element, tuple):
+ if self.unpack and isinstance(element, tuple) and len(element) > 1:
# If new data is a (id, data, tag) tuple use that except for multi-action "hstack" merges which produce tuples
if isinstance(data, tuple) and (len(self.action) <= 1 or self.merge != "hstack"):
return data
@@ -314,6 +318,9 @@ class Task:
action outputs
"""
+ # Log inputs
+ logger.debug("Inputs: %s", inputs)
+
# Execute action and get outputs
outputs = action(inputs)
@@ -321,6 +328,9 @@ class Task:
if isinstance(outputs, types.GeneratorType):
outputs = list(outputs)
+ # Log outputs
+ logger.debug("Outputs: %s", outputs)
+
return outputs
def postprocess(self, outputs):
diff --git a/src/python/txtai/workflow/task/template.py b/src/python/txtai/workflow/task/template.py
new file mode 100644
index 00000000..dbe25324
--- /dev/null
+++ b/src/python/txtai/workflow/task/template.py
@@ -0,0 +1,121 @@
+"""
+Template module
+"""
+
+from string import Formatter
+
+from .file import Task
+
+
+class TemplateTask(Task):
+ """
+ Task that generates text from a template and task inputs. Templates can be used to prepare data for a number of tasks
+ including generating large languge model (LLM) prompts.
+ """
+
+ def register(self, template=None, rules=None, strict=True):
+ """
+ Read template parameters.
+
+ Args:
+ template: prompt template
+ rules: parameter rules
+ strict: requires all task inputs to be consumed by template, defaults to True
+ """
+
+ # pylint: disable=W0201
+ # Template text
+ self.template = template if template else self.defaulttemplate()
+
+ # Template processing rules
+ self.rules = rules if rules else self.defaultrules()
+
+ # Create formatter
+ self.formatter = TemplateFormatter() if strict else Formatter()
+
+ def prepare(self, element):
+ # Check if element matches any processing rules
+ match = self.match(element)
+ if match:
+ return match
+
+ # Apply template processing, if applicable
+ if self.template:
+ # Pass dictionary as named prompt template parameters
+ if isinstance(element, dict):
+ return self.formatter.format(self.template, **element)
+
+ # Pass tuple as prompt template parameters (arg0 - argN)
+ if isinstance(element, tuple):
+ return self.formatter.format(self.template, **{f"arg{i}": x for i, x in enumerate(element)})
+
+ # Default behavior is to use input as {text} parameter in prompt template
+ return self.formatter.format(self.template, text=element)
+
+ # Return original inputs when no prompt provided
+ return element
+
+ def defaulttemplate(self):
+ """
+ Generates a default template for this task. Base method returns None.
+
+ Returns:
+ default template
+ """
+
+ return None
+
+ def defaultrules(self):
+ """
+ Generates a default rules for this task. Base method returns an empty dictionary.
+
+ Returns:
+ default rules
+ """
+
+ return {}
+
+ def match(self, element):
+ """
+ Check if element matches any processing rules.
+
+ Args:
+ element: input element
+
+ Returns:
+ matching value if found, None otherwise
+ """
+
+ if self.rules and isinstance(element, dict):
+ # Check if any rules are matched
+ for key, value in self.rules.items():
+ if element[key] == value:
+ return element[key]
+
+ return None
+
+
+class ExtractorTask(TemplateTask):
+ """
+ Template task that prepares input for an extractor pipeline.
+ """
+
+ def prepare(self, element):
+ # Allow partial input with the "question" field used to complete prompt template
+ if isinstance(element, dict):
+ element["question"] = super().prepare(element["question"])
+ return element
+
+ # Default mode is to use element text for both query and question
+ return {"query": element, "question": super().prepare(element)}
+
+
+class TemplateFormatter(Formatter):
+ """
+ Helper class used to format template checks.
+ """
+
+ def check_unused_args(self, used_args, args, kwargs):
+ difference = set(kwargs).difference(used_args)
+ if difference:
+ raise KeyError(difference)