Improve JSON extraction, add extraction to Task Memory

CHANGELOG.md

@@ -9,6 +9,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Added
 - `AstraDbVectorStoreDriver` to support DataStax Astra DB as a vector store.
 - Ability to set custom schema properties on Tool Activities via `extra_schema_properties`.
+- `extract_json` and `extract_csv` Activities to `TaskMemoryClient`.
+- `extract_json_namespace` and `extract_csv_namespace` methods to `TaskMemory`.
+
+### Changed
+- **BREAKING**: Split parameter `JsonExtractionEngine.template_generator` into `system_template_generator` and `user_template_generator`.
+- **BREAKING**: Split parameter `CsvExtractionEngine.template_generator` into `system_template_generator` and `user_template_generator`.
+- **BREAKING**: Split `JsonExtractionEngine.extract` into `extract_text` and `extract_artifacts`.
+- **BREAKING**: Split `CsvExtractionEngine.extract` into `extract_text` and `extract_artifacts`.
+- Parse json from LLM output before loading in `JsonExtractionEngine`.
+
+### Fixed
+- Missing implementations of `csv_extraction_engine` and `json_extraction_engine` in `TextArtifactStorage`.
 
 ## [0.29.0] - 2024-07-30
 

docs/griptape-framework/engines/extraction-engines.md

@@ -32,7 +32,7 @@ Charlie is 40 and lives in Texas.
 """
 
 # Extract CSV rows using the engine
-result = csv_engine.extract(sample_text, column_names=["name", "age", "location"])
+result = csv_engine.extract_text(sample_text, column_names=["name", "age", "location"])
 
 for row in result.value:
     print(row.to_text())
@@ -73,7 +73,7 @@ user_schema = Schema(
 ).json_schema("UserSchema")
 
 # Extract data using the engine
-result = json_engine.extract(sample_json_text, template_schema=user_schema)
+result = json_engine.extract_text(sample_json_text, template_schema=user_schema)
 
 for artifact in result.value:
     print(artifact.value)

griptape/engines/extraction/base_extraction_engine.py

@@ -5,10 +5,11 @@
 
 from attrs import Attribute, Factory, define, field
 
+from griptape.artifacts import ListArtifact, TextArtifact
 from griptape.chunkers import BaseChunker, TextChunker
 
 if TYPE_CHECKING:
-    from griptape.artifacts import ErrorArtifact, ListArtifact
+    from griptape.artifacts import ErrorArtifact
     from griptape.drivers import BasePromptDriver
     from griptape.rules import Ruleset
 
@@ -45,10 +46,15 @@ def min_response_tokens(self) -> int:
         )
 
     @abstractmethod
-    def extract(
+    def extract_artifacts(
         self,
-        text: str | ListArtifact,
+        artifacts: ListArtifact,
         *,
         rulesets: Optional[list[Ruleset]] = None,
         **kwargs,
     ) -> ListArtifact | ErrorArtifact: ...
+
+    def extract_text(
+        self, text: str, *, rulesets: Optional[list[Ruleset]] = None, **kwargs
+    ) -> ListArtifact | ErrorArtifact:
+        return self.extract_artifacts(ListArtifact([TextArtifact(text)]), rulesets=rulesets, **kwargs)

griptape/engines/extraction/csv_extraction_engine.py

@@ -18,11 +18,12 @@
 
 @define
 class CsvExtractionEngine(BaseExtractionEngine):
-    template_generator: J2 = field(default=Factory(lambda: J2("engines/extraction/csv_extraction.j2")), kw_only=True)
+    system_template_generator: J2 = field(default=Factory(lambda: J2("engines/csv_extraction/system.j2")), kw_only=True)
+    user_template_generator: J2 = field(default=Factory(lambda: J2("engines/csv_extraction/user.j2")), kw_only=True)
 
-    def extract(
+    def extract_artifacts(
         self,
-        text: str | ListArtifact,
+        artifacts: ListArtifact,
         *,
         rulesets: Optional[list[Ruleset]] = None,
         column_names: Optional[list[str]] = None,
@@ -33,7 +34,7 @@ def extract(
         try:
             return ListArtifact(
                 self._extract_rec(
-                    cast(list[TextArtifact], text.value) if isinstance(text, ListArtifact) else [TextArtifact(text)],
+                    cast(list[TextArtifact], artifacts.value),
                     column_names,
                     [],
                     rulesets=rulesets,
@@ -60,32 +61,49 @@ def _extract_rec(
         rulesets: Optional[list[Ruleset]] = None,
     ) -> list[CsvRowArtifact]:
         artifacts_text = self.chunk_joiner.join([a.value for a in artifacts])
-        full_text = self.template_generator.render(
+        system_prompt = self.system_template_generator.render(
             column_names=column_names,
-            text=artifacts_text,
             rulesets=J2("rulesets/rulesets.j2").render(rulesets=rulesets),
         )
+        user_prompt = self.user_template_generator.render(
+            text=artifacts_text,
+        )
 
-        if self.prompt_driver.tokenizer.count_input_tokens_left(full_text) >= self.min_response_tokens:
+        if (
+            self.prompt_driver.tokenizer.count_input_tokens_left(system_prompt + user_prompt)
+            >= self.min_response_tokens
+        ):
             rows.extend(
                 self.text_to_csv_rows(
-                    self.prompt_driver.run(PromptStack(messages=[Message(full_text, role=Message.USER_ROLE)])).value,
+                    self.prompt_driver.run(
+                        PromptStack(
+                            messages=[
+                                Message(system_prompt, role=Message.SYSTEM_ROLE),
+                                Message(user_prompt, role=Message.USER_ROLE),
+                            ]
+                        )
+                    ).value,
                     column_names,
                 ),
             )
 
             return rows
         else:
             chunks = self.chunker.chunk(artifacts_text)
-            partial_text = self.template_generator.render(
-                column_names=column_names,
+            partial_text = self.user_template_generator.render(
                 text=chunks[0].value,
-                rulesets=J2("rulesets/rulesets.j2").render(rulesets=rulesets),
             )
 
             rows.extend(
                 self.text_to_csv_rows(
-                    self.prompt_driver.run(PromptStack(messages=[Message(partial_text, role=Message.USER_ROLE)])).value,
+                    self.prompt_driver.run(
+                        PromptStack(
+                            messages=[
+                                Message(system_prompt, role=Message.SYSTEM_ROLE),
+                                Message(partial_text, role=Message.USER_ROLE),
+                            ]
+                        )
+                    ).value,
                     column_names,
                 ),
             )

griptape/engines/extraction/json_extraction_engine.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import re
 from typing import TYPE_CHECKING, Optional, cast
 
 from attrs import Factory, define, field
@@ -17,11 +18,16 @@
 
 @define
 class JsonExtractionEngine(BaseExtractionEngine):
-    template_generator: J2 = field(default=Factory(lambda: J2("engines/extraction/json_extraction.j2")), kw_only=True)
+    JSON_PATTERN = r"(?s)[^\[]*(\[.*\])"
 
-    def extract(
+    system_template_generator: J2 = field(
+        default=Factory(lambda: J2("engines/json_extraction/system.j2")), kw_only=True
+    )
+    user_template_generator: J2 = field(default=Factory(lambda: J2("engines/json_extraction/user.j2")), kw_only=True)
+
+    def extract_artifacts(
         self,
-        text: str | ListArtifact,
+        artifacts: ListArtifact,
         *,
         rulesets: Optional[list[Ruleset]] = None,
         template_schema: Optional[list[dict]] = None,
@@ -34,7 +40,7 @@ def extract(
 
             return ListArtifact(
                 self._extract_rec(
-                    cast(list[TextArtifact], text.value) if isinstance(text, ListArtifact) else [TextArtifact(text)],
+                    cast(list[TextArtifact], artifacts.value),
                     json_schema,
                     [],
                     rulesets=rulesets,
@@ -45,7 +51,12 @@ def extract(
             return ErrorArtifact(f"error extracting JSON: {e}")
 
     def json_to_text_artifacts(self, json_input: str) -> list[TextArtifact]:
-        return [TextArtifact(json.dumps(e)) for e in json.loads(json_input)]
+        json_matches = re.findall(self.JSON_PATTERN, json_input, re.DOTALL)
+
+        if json_matches:
+            return [TextArtifact(json.dumps(e)) for e in json.loads(json_matches[-1])]
+        else:
+            return []
 
     def _extract_rec(
         self,
@@ -55,31 +66,48 @@ def _extract_rec(
         rulesets: Optional[list[Ruleset]] = None,
     ) -> list[TextArtifact]:
         artifacts_text = self.chunk_joiner.join([a.value for a in artifacts])
-        full_text = self.template_generator.render(
+        system_prompt = self.system_template_generator.render(
             json_template_schema=json_template_schema,
-            text=artifacts_text,
             rulesets=J2("rulesets/rulesets.j2").render(rulesets=rulesets),
         )
+        user_prompt = self.user_template_generator.render(
+            text=artifacts_text,
+        )
 
-        if self.prompt_driver.tokenizer.count_input_tokens_left(full_text) >= self.min_response_tokens:
+        if (
+            self.prompt_driver.tokenizer.count_input_tokens_left(user_prompt + system_prompt)
+            >= self.min_response_tokens
+        ):
             extractions.extend(
                 self.json_to_text_artifacts(
-                    self.prompt_driver.run(PromptStack(messages=[Message(full_text, role=Message.USER_ROLE)])).value,
+                    self.prompt_driver.run(
+                        PromptStack(
+                            messages=[
+                                Message(system_prompt, role=Message.SYSTEM_ROLE),
+                                Message(user_prompt, role=Message.USER_ROLE),
+                            ]
+                        )
+                    ).value
                 ),
             )
 
             return extractions
         else:
             chunks = self.chunker.chunk(artifacts_text)
-            partial_text = self.template_generator.render(
-                template_schema=json_template_schema,
+            partial_text = self.user_template_generator.render(
                 text=chunks[0].value,
-                rulesets=J2("rulesets/rulesets.j2").render(rulesets=rulesets),
             )
 
             extractions.extend(
                 self.json_to_text_artifacts(
-                    self.prompt_driver.run(PromptStack(messages=[Message(partial_text, role=Message.USER_ROLE)])).value,
+                    self.prompt_driver.run(
+                        PromptStack(
+                            messages=[
+                                Message(system_prompt, role=Message.SYSTEM_ROLE),
+                                Message(partial_text, role=Message.USER_ROLE),
+                            ]
+                        )
+                    ).value,
                 ),
             )
 

griptape/memory/task/storage/base_artifact_storage.py

@@ -6,7 +6,7 @@
 from attrs import define
 
 if TYPE_CHECKING:
-    from griptape.artifacts import BaseArtifact, InfoArtifact, ListArtifact, TextArtifact
+    from griptape.artifacts import BaseArtifact, ErrorArtifact, InfoArtifact, ListArtifact, TextArtifact
 
 
 @define
@@ -25,3 +25,9 @@ def summarize(self, namespace: str) -> TextArtifact | InfoArtifact: ...
 
     @abstractmethod
     def query(self, namespace: str, query: str, metadata: Any = None) -> BaseArtifact: ...
+
+    @abstractmethod
+    def extract_csv(self, namespace: str) -> ListArtifact | InfoArtifact | ErrorArtifact: ...
+
+    @abstractmethod
+    def extract_json(self, namespace: str) -> ListArtifact | InfoArtifact | ErrorArtifact: ...

griptape/memory/task/storage/blob_artifact_storage.py

@@ -32,3 +32,9 @@ def summarize(self, namespace: str) -> InfoArtifact:
 
     def query(self, namespace: str, query: str, metadata: Any = None) -> BaseArtifact:
         return InfoArtifact("can't query artifacts")
+
+    def extract_csv(self, namespace: str) -> InfoArtifact:
+        return InfoArtifact("can't extract csv")
+
+    def extract_json(self, namespace: str) -> InfoArtifact:
+        return InfoArtifact("can't extract json")

griptape/memory/task/storage/text_artifact_storage.py

@@ -9,6 +9,7 @@
 from griptape.memory.task.storage import BaseArtifactStorage
 
 if TYPE_CHECKING:
+    from griptape.artifacts import ErrorArtifact
     from griptape.drivers import BaseVectorStoreDriver
     from griptape.engines import BaseSummaryEngine, CsvExtractionEngine, JsonExtractionEngine
 
@@ -45,6 +46,18 @@ def summarize(self, namespace: str) -> TextArtifact:
 
         return self.summary_engine.summarize_artifacts(self.load_artifacts(namespace))
 
+    def extract_csv(self, namespace: str) -> ListArtifact | ErrorArtifact:
+        if self.csv_extraction_engine is None:
+            raise ValueError("Csv extraction engine is not set.")
+
+        return self.csv_extraction_engine.extract_artifacts(self.load_artifacts(namespace))
+
+    def extract_json(self, namespace: str) -> ListArtifact | ErrorArtifact:
+        if self.json_extraction_engine is None:
+            raise ValueError("Json extraction engine is not set.")
+
+        return self.json_extraction_engine.extract_artifacts(self.load_artifacts(namespace))
+
     def query(self, namespace: str, query: str, metadata: Any = None) -> BaseArtifact:
         if self.rag_engine is None:
             raise ValueError("rag_engine is not set")

griptape/memory/task/task_memory.py

@@ -139,3 +139,19 @@ def query_namespace(self, namespace: str, query: str) -> BaseArtifact:
             return storage.query(namespace=namespace, query=query, metadata=self.namespace_metadata.get(namespace))
         else:
             return InfoArtifact("Can't find memory content")
+
+    def extract_json_namespace(self, namespace: str) -> ListArtifact | InfoArtifact | ErrorArtifact:
+        storage = self.namespace_storage.get(namespace)
+
+        if storage:
+            return storage.extract_json(namespace)
+        else:
+            return ErrorArtifact("Can't find memory content")
+
+    def extract_csv_namespace(self, namespace: str) -> ListArtifact | InfoArtifact | ErrorArtifact:
+        storage = self.namespace_storage.get(namespace)
+
+        if storage:
+            return storage.extract_csv(namespace)
+        else:
+            return ErrorArtifact("Can't find memory content")

griptape/tasks/extraction_task.py

@@ -21,4 +21,4 @@ def extraction_engine(self) -> BaseExtractionEngine:
         return self._extraction_engine
 
     def run(self) -> ListArtifact | ErrorArtifact:
-        return self.extraction_engine.extract(self.input.to_text(), rulesets=self.all_rulesets, **self.args)
+        return self.extraction_engine.extract_text(self.input.to_text(), rulesets=self.all_rulesets, **self.args)

griptape/templates/engines/csv_extraction/system.j2

@@ -0,0 +1,7 @@
+Don't add the header row. Don't use markdown formatting for output. Fields containing line breaks (CRLF), double quotes, and commas should be enclosed in double-quotes.
+Column Names: """{{ column_names }}"""
+
+{% if rulesets %}
+
+{{ rulesets }}
+{% endif %}

griptape/templates/engines/csv_extraction/user.j2

@@ -0,0 +1,4 @@
+Extract information from the Text based on the Column Names and output it as a CSV file.
+Text: """{{ text }}"""
+
+Answer:

griptape/templates/engines/json_extraction/system.j2

@@ -0,0 +1,6 @@
+Extraction Template JSON Schema: """{{ json_template_schema }}"""
+
+{% if rulesets %}
+
+{{ rulesets }}
+{% endif %}

griptape/templates/engines/extraction/json_extraction.j2 → griptape/templates/engines/json_extraction/user.j2

@@ -1,11 +1,4 @@
-Text: """{{ text }}"""
-
-Extraction Template JSON Schema: """{{ json_template_schema }}"""
-
 Extract information from the Text based on the Extraction Template JSON Schema into an array of JSON objects.
-{% if rulesets %}
-
-{{ rulesets }}
-{% endif %}
+Text: """{{ text }}"""
 
 JSON array: