Yusaki/MiroFish
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Adjust relationship prompt to infer connections from character traits

Browse source
This commit is contained in:
_Yusaki 2026-03-13 19:34:51 +07:00
parent 10a85e76d6
commit 49a86c622a
1 changed files with 12 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
backend/app/services/llm_graph_builder.py
View file

@ -36,18 +36,21 @@ ENTITY_EXTRACT_PROMPT = (
)
RELATIONSHIP_EXTRACT_PROMPT = (
"You are a knowledge graph relationship extraction engine. Given a text section, a list of known entities, "
"and an ontology schema, extract all relationships between the entities.\n\n"
"ONTOLOGY SCHEMA (edge types):\n%s\n\n"
"You are a knowledge graph relationship inference engine. Given a text section describing characters/entities "
"in a social simulation, a list of known entities, and relationship types, infer all plausible relationships.\n\n"
"RELATIONSHIP TYPES:\n%s\n\n"
"KNOWN ENTITIES:\n%s\n\n"
"RULES:\n"
"1. Find relationships between the known entities that match the edge_types defined above.\n"
"2. Each relationship needs: name (must match an edge_type name), source (entity name), target (entity name), "
"fact (1 sentence describing the specific relationship found in the text).\n"
"1. Infer relationships between the known entities based on their described traits, roles, goals, and behaviors.\n"
"2. Each relationship needs: name (must match a relationship type name above), source (entity name), target (entity name), "
"fact (1 sentence explaining WHY this relationship is likely based on the text).\n"
"3. Both source and target MUST be from the known entities list.\n"
"4. Only extract relationships explicitly stated or strongly implied in the text.\n"
"5. Extract ALL relationships you can find — be thorough.\n"
"6. If no relationships are found, return an empty array.\n\n"
"4. Include both explicit relationships AND strongly implied ones based on complementary/conflicting traits.\n"
" Examples of inference: if player A 'trades supplies' and player B 'buys tools', infer TRADE_WITH.\n"
" If player A 'mediates conflicts' and player B 'causes chaos', infer OPPOSES.\n"
" If two players share goals or would naturally work together, infer COLLABORATES_WITH.\n"
"5. Be thorough — extract as many plausible relationships as the text supports.\n"
"6. Do NOT invent relationships with no textual basis. Each fact must reference something from the text.\n\n"
'Return JSON: a single key "relationships" with an array of objects, each having keys: name, source, target, fact.'
)