semantic_analyzer verschachtelete Strukturen
This commit is contained in:
parent
7fc316d284
commit
d93b9b30ae
|
|
@ -1,7 +1,6 @@
|
|||
"""
|
||||
app/services/semantic_analyzer.py
|
||||
Zweck: Asynchroner Service zur Zuweisung von Kanten zu Text-Chunks mittels LLM.
|
||||
Nutzt Templates aus prompts.yaml.
|
||||
app/services/semantic_analyzer.py — Edge Validation & Filtering
|
||||
Version: 1.1 (Robust JSON Parsing)
|
||||
"""
|
||||
|
||||
import json
|
||||
|
|
@ -28,17 +27,22 @@ class SemanticAnalyzer:
|
|||
|
||||
# 1. Prompt laden
|
||||
prompt_template = self.llm.prompts.get("edge_allocation_template")
|
||||
|
||||
# Fallback, falls Prompt nicht in YAML definiert ist (für Tests ohne volle Config)
|
||||
if not prompt_template:
|
||||
logger.error("Prompt 'edge_allocation_template' in prompts.yaml nicht gefunden.")
|
||||
return []
|
||||
prompt_template = (
|
||||
"TASK: Wähle aus den Kandidaten die relevanten Kanten für den Text.\n"
|
||||
"TEXT: {chunk_text}\n"
|
||||
"KANDIDATEN: {edge_list}\n"
|
||||
"OUTPUT: JSON Liste von Strings [\"kind:target\"]."
|
||||
)
|
||||
|
||||
# 2. Kandidaten-Liste formatieren
|
||||
# Wir übergeben die Kanten als einfache Liste, damit das LLM sie auswählen kann.
|
||||
edges_str = "\n".join([f"- {e}" for e in all_edges])
|
||||
|
||||
# 3. Prompt füllen
|
||||
final_prompt = prompt_template.format(
|
||||
chunk_text=chunk_text[:3000], # Truncate safety
|
||||
chunk_text=chunk_text[:3000],
|
||||
edge_list=edges_str
|
||||
)
|
||||
|
||||
|
|
@ -49,32 +53,41 @@ class SemanticAnalyzer:
|
|||
force_json=True
|
||||
)
|
||||
|
||||
# 5. Parsing
|
||||
# 5. Parsing & Cleaning
|
||||
clean_json = response_json.replace("```json", "").replace("```", "").strip()
|
||||
|
||||
# Fallback für leere Antworten
|
||||
if not clean_json:
|
||||
return []
|
||||
if not clean_json: return []
|
||||
|
||||
data = json.loads(clean_json)
|
||||
valid_edges = []
|
||||
|
||||
# 6. Validierung: Wir erwarten eine Liste von Strings
|
||||
# 6. Robuste Validierung (List vs Dict)
|
||||
if isinstance(data, list):
|
||||
# Filtern: Nur Strings zurückgeben, die auch in der Input-Liste waren (Sicherheit)
|
||||
# oder zumindest das korrekte Format haben.
|
||||
# Standardfall: ["kind:target", ...]
|
||||
valid_edges = [str(e) for e in data if isinstance(e, str) and ":" in e]
|
||||
return valid_edges
|
||||
elif isinstance(data, dict):
|
||||
# Manchmal packt das LLM es in {"edges": [...]}
|
||||
for key, val in data.items():
|
||||
if isinstance(val, list):
|
||||
return [str(e) for e in val if isinstance(e, str)]
|
||||
|
||||
logger.warning(f"SemanticAnalyzer: Unerwartetes JSON Format: {str(data)[:100]}")
|
||||
return []
|
||||
elif isinstance(data, dict):
|
||||
# Abweichende Formate behandeln
|
||||
for key, val in data.items():
|
||||
# Fall A: {"edges": ["kind:target"]}
|
||||
if key.lower() in ["edges", "results", "kanten"] and isinstance(val, list):
|
||||
valid_edges.extend([str(e) for e in val if isinstance(e, str) and ":" in e])
|
||||
|
||||
# Fall B: {"kind": "target"} (Das beobachtete Format im Log)
|
||||
elif isinstance(val, str):
|
||||
# Wir rekonstruieren "kind:target"
|
||||
valid_edges.append(f"{key}:{val}")
|
||||
|
||||
# Fall C: {"kind": ["target1", "target2"]}
|
||||
elif isinstance(val, list):
|
||||
for target in val:
|
||||
if isinstance(target, str):
|
||||
valid_edges.append(f"{key}:{target}")
|
||||
|
||||
# Safety: Filtere nur Kanten, die halbwegs valide aussehen
|
||||
return [e for e in valid_edges if ":" in e]
|
||||
|
||||
except json.JSONDecodeError:
|
||||
logger.warning("SemanticAnalyzer: LLM lieferte kein valides JSON. Keine Kanten zugewiesen.")
|
||||
logger.warning("SemanticAnalyzer: LLM lieferte kein valides JSON. Ignoriere Zuweisung.")
|
||||
return []
|
||||
except Exception as e:
|
||||
logger.error(f"SemanticAnalyzer Error: {e}")
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user