Enhanced the edge deduplication prompts to better recognize semantically
equivalent facts that use different phrasings:
- Self-referential relationships ("X is a sub-agency of X" = "X is its own sub-agency")
- Active vs passive voice ("A awarded contract to B" = "B received contract from A")
- Numeric format equivalence ($1M = $1,000,000)
- Entity aliases (DoD = Department of Defense)
Added integration tests that verify the LLM correctly identifies semantic
duplicates with the improved prompts.
178 lines
6.4 KiB
Python
178 lines
6.4 KiB
Python
"""
|
|
Copyright 2024, Zep Software, Inc.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
"""
|
|
|
|
from typing import Any, Protocol, TypedDict
|
|
|
|
from pydantic import BaseModel, Field
|
|
|
|
from .models import Message, PromptFunction, PromptVersion
|
|
from .prompt_helpers import to_prompt_json
|
|
|
|
|
|
class EdgeDuplicate(BaseModel):
|
|
duplicate_facts: list[int] = Field(
|
|
...,
|
|
description='List of idx values of any duplicate facts. If no duplicate facts are found, default to empty list.',
|
|
)
|
|
contradicted_facts: list[int] = Field(
|
|
...,
|
|
description='List of idx values of facts that should be invalidated. If no facts should be invalidated, the list should be empty.',
|
|
)
|
|
fact_type: str = Field(..., description='One of the provided fact types or DEFAULT')
|
|
|
|
|
|
class UniqueFact(BaseModel):
|
|
uuid: str = Field(..., description='unique identifier of the fact')
|
|
fact: str = Field(..., description='fact of a unique edge')
|
|
|
|
|
|
class UniqueFacts(BaseModel):
|
|
unique_facts: list[UniqueFact]
|
|
|
|
|
|
class Prompt(Protocol):
|
|
edge: PromptVersion
|
|
edge_list: PromptVersion
|
|
resolve_edge: PromptVersion
|
|
|
|
|
|
class Versions(TypedDict):
|
|
edge: PromptFunction
|
|
edge_list: PromptFunction
|
|
resolve_edge: PromptFunction
|
|
|
|
|
|
def edge(context: dict[str, Any]) -> list[Message]:
|
|
return [
|
|
Message(
|
|
role='system',
|
|
content='You are a helpful assistant that de-duplicates edges from edge lists.',
|
|
),
|
|
Message(
|
|
role='user',
|
|
content=f"""
|
|
Given the following context, determine whether the New Edge represents any of the edges in the list of Existing Edges.
|
|
|
|
<EXISTING EDGES>
|
|
{to_prompt_json(context['related_edges'])}
|
|
</EXISTING EDGES>
|
|
|
|
<NEW EDGE>
|
|
{to_prompt_json(context['extracted_edges'])}
|
|
</NEW EDGE>
|
|
|
|
Task:
|
|
If the New Edges represents the same factual information as any edge in Existing Edges, return the id of the duplicate fact
|
|
as part of the list of duplicate_facts.
|
|
If the NEW EDGE is not a duplicate of any of the EXISTING EDGES, return an empty list.
|
|
|
|
Guidelines:
|
|
1. The facts do not need to be completely identical to be duplicates, they just need to express the same information.
|
|
""",
|
|
),
|
|
]
|
|
|
|
|
|
def edge_list(context: dict[str, Any]) -> list[Message]:
|
|
return [
|
|
Message(
|
|
role='system',
|
|
content='You are a helpful assistant that de-duplicates edges from edge lists.',
|
|
),
|
|
Message(
|
|
role='user',
|
|
content=f"""
|
|
Given the following context, find all of the duplicates in a list of facts:
|
|
|
|
Facts:
|
|
{to_prompt_json(context['edges'])}
|
|
|
|
Task:
|
|
If any facts in Facts is a duplicate of another fact, return a new fact with one of their uuid's.
|
|
|
|
Guidelines:
|
|
1. identical or near identical facts are duplicates
|
|
2. Facts are also duplicates if they are represented by similar sentences
|
|
3. Facts will often discuss the same or similar relation between identical entities
|
|
4. The final list should have only unique facts. If 3 facts are all duplicates of each other, only one of their
|
|
facts should be in the response
|
|
""",
|
|
),
|
|
]
|
|
|
|
|
|
def resolve_edge(context: dict[str, Any]) -> list[Message]:
|
|
return [
|
|
Message(
|
|
role='system',
|
|
content='You are a helpful assistant that de-duplicates facts from fact lists and determines which existing '
|
|
'facts are contradicted by the new fact.',
|
|
),
|
|
Message(
|
|
role='user',
|
|
content=f"""
|
|
Task:
|
|
You will receive TWO separate lists of facts. Each list uses 'idx' as its index field, starting from 0.
|
|
|
|
1. DUPLICATE DETECTION:
|
|
- If the NEW FACT represents identical factual information as any fact in EXISTING FACTS, return those idx values in duplicate_facts.
|
|
- Facts with similar information that contain key differences should NOT be marked as duplicates.
|
|
- Facts ARE duplicates if they convey the same information using different wording.
|
|
- Return idx values from EXISTING FACTS.
|
|
- If no duplicates, return an empty list for duplicate_facts.
|
|
|
|
2. FACT TYPE CLASSIFICATION:
|
|
- Given the predefined FACT TYPES, determine if the NEW FACT should be classified as one of these types.
|
|
- Return the fact type as fact_type or DEFAULT if NEW FACT is not one of the FACT TYPES.
|
|
|
|
3. CONTRADICTION DETECTION:
|
|
- Based on FACT INVALIDATION CANDIDATES and NEW FACT, determine which facts the new fact contradicts.
|
|
- Return idx values from FACT INVALIDATION CANDIDATES.
|
|
- If no contradictions, return an empty list for contradicted_facts.
|
|
|
|
IMPORTANT:
|
|
- duplicate_facts: Use ONLY 'idx' values from EXISTING FACTS
|
|
- contradicted_facts: Use ONLY 'idx' values from FACT INVALIDATION CANDIDATES
|
|
- These are two separate lists with independent idx ranges starting from 0
|
|
|
|
Guidelines:
|
|
1. Some facts may be very similar but will have key differences, particularly around numeric values in the facts.
|
|
Do not mark these facts as duplicates.
|
|
2. Same values in different formats ARE duplicates (e.g., "$1M" = "$1,000,000").
|
|
3. Active/passive voice variations ARE duplicates (e.g., "X awarded Y" = "Y received from X").
|
|
4. Self-referential rewordings ARE duplicates (e.g., "X is its own Y" = "X is a Y of X").
|
|
|
|
<FACT TYPES>
|
|
{context['edge_types']}
|
|
</FACT TYPES>
|
|
|
|
<EXISTING FACTS>
|
|
{context['existing_edges']}
|
|
</EXISTING FACTS>
|
|
|
|
<FACT INVALIDATION CANDIDATES>
|
|
{context['edge_invalidation_candidates']}
|
|
</FACT INVALIDATION CANDIDATES>
|
|
|
|
<NEW FACT>
|
|
{context['new_edge']}
|
|
</NEW FACT>
|
|
""",
|
|
),
|
|
]
|
|
|
|
|
|
versions: Versions = {'edge': edge, 'edge_list': edge_list, 'resolve_edge': resolve_edge}
|