gmakstutis/graphiti
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Optimize edge deduplication prompt for caching and clarity (#970)

Browse source 
* Optimize edge deduplication prompt for caching and clarity

- Restructure prompt to place invariant instructions at top and dynamic context at bottom for better LLM caching
- Change 'id' to 'idx' in edge context lists to avoid confusion with other identifiers
- Remove 'fact_type_id' from edge types context as LLM only needs fact_type_name
- Remove dynamic range values from prompt instructions (e.g., "range 0-N")
- Add debug logging before LLM call to track input sizes
- Add validation logging after LLM response to catch invalid idx values
- Clarify that duplicate_facts uses EXISTING FACTS idx and contradicted_facts uses INVALIDATION CANDIDATES idx

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* Address terminology consistency and edge case logging

- Update Pydantic field descriptions to use 'idx' instead of 'ids' for consistency
- Fix debug logging to handle empty list edge case (avoid 'idx 0--1' display)

Note on review feedback:
- Validation is intentionally non-redundant: warnings provide visibility, list comprehensions ensure robustness
- WARNING level is appropriate for LLM output issues (not system errors)
- Existing test coverage is sufficient for this defensive logging addition

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Daniel Chalef 2025-10-02 17:07:43 -07:00 • committed by GitHub
parent b28bd92c16
commit 4a307dbf10
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 69 additions and 32 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

67
graphiti_core/prompts/dedupe_edges.py
View file

@ -25,11 +25,11 @@ from .prompt_helpers import to_prompt_json
class EdgeDuplicate(BaseModel):
duplicate_facts: list[int] = Field(
...,
description='List of ids of any duplicate facts. If no duplicate facts are found, default to empty list.',
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 ids of facts that should be invalidated. If no facts should be invalidated, the list should be empty.',
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')
@ -124,37 +124,48 @@ def resolve_edge(context: dict[str, Any]) -> list[Message]:
Message(
role='user',
content=f"""
<NEW FACT>
{context['new_edge']}
</NEW FACT>
<EXISTING FACTS>
{context['existing_edges']}
</EXISTING FACTS>
<FACT INVALIDATION CANDIDATES>
{context['edge_invalidation_candidates']}
</FACT INVALIDATION CANDIDATES>
<FACT TYPES>
{context['edge_types']}
</FACT TYPES>
Task:
If the NEW FACT represents identical factual information of one or more in EXISTING FACTS, return the idx of the duplicate facts.
Facts with similar information that contain key differences should not be marked as duplicates.
If the NEW FACT is not a duplicate of any of the EXISTING FACTS, return an empty list.
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.
Based on the provided FACT INVALIDATION CANDIDATES and NEW FACT, determine which existing facts the new fact contradicts.
Return a list containing all idx's of the facts that are contradicted by the NEW FACT.
If there are no contradicted facts, return an empty list.
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.
- 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.
<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>
""",
),
]

34
graphiti_core/utils/maintenance/edge_operations.py
View file

@ -475,20 +475,19 @@ async def resolve_extracted_edge(
start = time()
# Prepare context for LLM
related_edges_context = [{'id': i, 'fact': edge.fact} for i, edge in enumerate(related_edges)]
related_edges_context = [{'idx': i, 'fact': edge.fact} for i, edge in enumerate(related_edges)]
invalidation_edge_candidates_context = [
{'id': i, 'fact': existing_edge.fact} for i, existing_edge in enumerate(existing_edges)
{'idx': i, 'fact': existing_edge.fact} for i, existing_edge in enumerate(existing_edges)
]
edge_types_context = (
[
{
'fact_type_id': i,
'fact_type_name': type_name,
'fact_type_description': type_model.__doc__,
}
for i, (type_name, type_model) in enumerate(edge_type_candidates.items())
for type_name, type_model in edge_type_candidates.items()
]
if edge_type_candidates is not None
else []
@ -501,6 +500,15 @@ async def resolve_extracted_edge(
'edge_types': edge_types_context,
}
if related_edges or existing_edges:
logger.debug(
'Resolving edge: sent %d EXISTING FACTS%s and %d INVALIDATION CANDIDATES%s',
len(related_edges),
f' (idx 0-{len(related_edges) - 1})' if related_edges else '',
len(existing_edges),
f' (idx 0-{len(existing_edges) - 1})' if existing_edges else '',
)
llm_response = await llm_client.generate_response(
prompt_library.dedupe_edges.resolve_edge(context),
response_model=EdgeDuplicate,
@ -509,6 +517,15 @@ async def resolve_extracted_edge(
response_object = EdgeDuplicate(**llm_response)
duplicate_facts = response_object.duplicate_facts
# Validate duplicate_facts are in valid range for EXISTING FACTS
invalid_duplicates = [i for i in duplicate_facts if i < 0 or i >= len(related_edges)]
if invalid_duplicates:
logger.warning(
'LLM returned invalid duplicate_facts idx values %s (valid range: 0-%d for EXISTING FACTS)',
invalid_duplicates,
len(related_edges) - 1,
)
duplicate_fact_ids: list[int] = [i for i in duplicate_facts if 0 <= i < len(related_edges)]
resolved_edge = extracted_edge
@ -521,6 +538,15 @@ async def resolve_extracted_edge(
contradicted_facts: list[int] = response_object.contradicted_facts
# Validate contradicted_facts are in valid range for INVALIDATION CANDIDATES
invalid_contradictions = [i for i in contradicted_facts if i < 0 or i >= len(existing_edges)]
if invalid_contradictions:
logger.warning(
'LLM returned invalid contradicted_facts idx values %s (valid range: 0-%d for INVALIDATION CANDIDATES)',
invalid_contradictions,
len(existing_edges) - 1,
)
invalidation_candidates: list[EntityEdge] = [
existing_edges[i] for i in contradicted_facts if 0 <= i < len(existing_edges)
]