Files
egutierrez 009d387d9a feat(0013): add paste_extract enricher (preview-only)
Modo preview puro — no escribe a operations.db. Recibe texto via
params.text y devuelve JSON con entidades y relaciones propuestas.
Cascada: extract_iocs (regex) siempre + extract_graph_hybrid
(GLiNER+GLiREL) opcional con use_hybrid=true. La aplicacion procesa
el JSON y persiste con dedupe via codigo C++ (extract_panel_apply).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-04 14:24:16 +02:00

304 lines
10 KiB
Python

#!/usr/bin/env python3
"""Enricher paste_extract — modo preview puro para el panel "Paste & Extract".
A diferencia del resto de enrichers, este NO escribe a operations.db. Recibe
el texto via `params.text` y devuelve un JSON con las entidades y relaciones
propuestas. La aplicacion (panel C++) procesa la propuesta, el usuario marca
cuales aceptar y la propia app persiste con dedupe via entity_ops.
Cascada de extraccion (graceful fallback):
1. `extract_iocs(text, types_list)` — regex puro, siempre disponible.
2. Si `use_hybrid=true` y los modelos cargan correctamente → ademas
`extract_graph_hybrid` (GLiNER + GLiREL + LLM opcional). Si la
carga del modelo falla por dependencias o por tiempo se ignora
silenciosamente — el panel muestra solo lo que llego.
Wire protocol estandar (issue 0026), pero tolera ausencia de `node_id` y de
`ops_db_path`: este enricher es global, no pertenece a un nodo.
Output JSON:
{
"entities": [
{"id": "<temp>", "type_ref": "Email", "name": "x@y.z",
"metadata": {...}, "source": "regex|gliner|llm",
"start": 12, "end": 27, "confidence": 1.0}
],
"relations": [
{"from_id": "<temp>", "to_id": "<temp>", "name": "RELATED_TO",
"source": "glirel|llm", "confidence": 0.7}
],
"stats": {"layers": ["regex"], "n_entities": N, "n_relations": M}
}
"""
from __future__ import annotations
import json
import os
import sys
_TYPE_MAP = {
"email": ("Email", "address"),
"ip_address": ("IPAddress", "address"),
"domain": ("Domain", "name"),
"file_hash": ("FileHash", "value"),
"crypto_wallet": ("CryptoWallet", "address"),
"cve_id": ("CVE", "id"),
"mac_address": ("MACAddress", "address"),
"phone_number": ("Phone", "number"),
}
def progress(p: float, stage: str = "") -> None:
sys.stderr.write(f"PROGRESS:{p:.2f} {stage}\n")
sys.stderr.flush()
def log(msg: str) -> None:
sys.stderr.write(f"{msg}\n")
sys.stderr.flush()
def _setup_registry_path(registry_root: str) -> None:
"""Intenta hacer importables los paquetes del registry.
Estrategia identica a otros enrichers:
1. `<this_dir>/_vendored/` — para distribuciones binarias.
2. `<registry_root>/python/functions/` — modo dev local.
"""
vendored = os.path.join(os.path.dirname(__file__), "_vendored")
if os.path.isdir(vendored):
if vendored not in sys.path:
sys.path.insert(0, vendored)
return
if registry_root:
py_funcs = os.path.join(registry_root, "python", "functions")
if os.path.isdir(py_funcs) and py_funcs not in sys.path:
sys.path.insert(0, py_funcs)
def _run_regex(text: str, types_list, max_entities: int) -> tuple[list, list]:
"""Capa 1: extract_iocs. Devuelve (entities, relations)."""
try:
from cybersecurity.extract_iocs import extract_iocs # type: ignore
except Exception as e:
log(f"extract_iocs no importable: {e}")
return [], []
iocs = extract_iocs(text, types_list)
seen = set()
entities = []
for i, it in enumerate(iocs):
ioc_type = it.get("type")
value = it.get("value") or it.get("address") or it.get("name") or ""
if not ioc_type or not value:
continue
key = (ioc_type, value)
if key in seen:
continue
seen.add(key)
type_ref, value_field = _TYPE_MAP.get(ioc_type, (ioc_type, "value"))
entities.append({
"id": f"tmp_e_{i}",
"type_ref": type_ref,
"name": value,
"metadata": {value_field: value},
"source": "regex",
"start": int(it.get("start", -1)),
"end": int(it.get("end", -1)),
"confidence": 1.0,
})
if len(entities) >= max_entities:
break
return entities, [] # regex no produce relaciones
def _try_run_hybrid(text: str, registry_root: str,
confidence_threshold: float = 0.6,
) -> tuple[list, list]:
"""Capa 2 opcional: extract_graph_hybrid. Devuelve (entities, relations).
Cualquier error o ausencia de dependencias se traga silenciosamente —
GLiNER/GLiREL pueden no estar instalados. La idea es no romper el
flow del panel cuando solo regex esta disponible.
"""
try:
# El pipeline esta en `pipelines/`, importarlo via path absoluto si
# el registry_root esta disponible. En modo vendored este pipeline
# quizas no este — fallback silencioso.
sys.path.insert(0, os.path.join(registry_root, "python", "functions",
"pipelines"))
from extract_graph_hybrid import extract_graph_hybrid # type: ignore
from datascience.gliner_load_model import gliner_load_model # type: ignore
from datascience.glirel_load_model import glirel_load_model # type: ignore
except Exception as e:
log(f"hybrid pipeline no disponible: {e}")
return [], []
# Schema generico para entidades semanticas. Si el caller quiere un
# schema custom, lo extendemos via params en una iteracion futura.
entity_schema = [
{"type_ref": "Person", "label": "person"},
{"type_ref": "Organization", "label": "organization"},
{"type_ref": "Location", "label": "location"},
]
relation_types = ["works_at", "located_in", "part_of", "related_to"]
try:
gliner_model = gliner_load_model()
glirel_model = glirel_load_model()
except Exception as e:
log(f"hybrid: load modelo fallo: {e}")
return [], []
try:
ents, rels = extract_graph_hybrid(
chunks=[text],
entity_schema=entity_schema,
relation_types=relation_types,
gliner_model=gliner_model,
glirel_model=glirel_model,
llm_chat_json=None,
confidence_threshold=confidence_threshold,
)
except Exception as e:
log(f"hybrid: extract fallo: {e}")
return [], []
out_entities = []
name_to_idx: dict[tuple[str, str], int] = {}
for i, ec in enumerate(ents):
idx = len(out_entities)
out_entities.append({
"id": f"tmp_h_{idx}",
"type_ref": ec.type_ref,
"name": ec.name,
"metadata": dict(getattr(ec, "attributes", {}) or {}),
"source": "hybrid",
"start": int((ec.attributes or {}).get("start", -1))
if hasattr(ec, "attributes") else -1,
"end": int((ec.attributes or {}).get("end", -1))
if hasattr(ec, "attributes") else -1,
"confidence": float(ec.confidence),
})
name_to_idx[(ec.type_ref, ec.name)] = idx
out_relations = []
for rc in rels:
# Mapear from/to (RelationCandidate) a tmp_id si los podemos casar.
from_key = (getattr(rc, "from_type_ref", None), getattr(rc, "from_name", None))
to_key = (getattr(rc, "to_type_ref", None), getattr(rc, "to_name", None))
if None in from_key or None in to_key:
continue
fi = name_to_idx.get(from_key)
ti = name_to_idx.get(to_key)
if fi is None or ti is None:
continue
out_relations.append({
"from_id": f"tmp_h_{fi}",
"to_id": f"tmp_h_{ti}",
"name": getattr(rc, "name", "RELATED_TO") or "RELATED_TO",
"source": "hybrid",
"confidence": float(getattr(rc, "confidence", 0.0)),
})
return out_entities, out_relations
def main() -> int:
raw = sys.stdin.read()
try:
ctx = json.loads(raw)
except Exception as e:
log(f"stdin not valid JSON: {e}")
return 2
params = ctx.get("params") or {}
registry_root = ctx.get("registry_root") or ""
text = (params.get("text") or "").strip()
if not text:
log("missing params.text")
print(json.dumps({"error": "missing params.text",
"entities": [], "relations": [],
"stats": {"layers": [], "n_entities": 0,
"n_relations": 0}}))
return 2
types_csv = (params.get("types") or "").strip()
types_list = [t.strip() for t in types_csv.split(",") if t.strip()] \
if types_csv else None
max_entities = int(params.get("max_entities", 200))
use_hybrid_raw = params.get("use_hybrid", False)
if isinstance(use_hybrid_raw, str):
use_hybrid = use_hybrid_raw.strip().lower() in ("1", "true", "yes", "on")
else:
use_hybrid = bool(use_hybrid_raw)
progress(0.05, "init")
_setup_registry_path(registry_root)
progress(0.20, "regex")
regex_entities, _ = _run_regex(text, types_list, max_entities)
layers = ["regex"]
hybrid_entities: list = []
hybrid_relations: list = []
if use_hybrid:
progress(0.40, "hybrid")
hybrid_entities, hybrid_relations = _try_run_hybrid(text, registry_root)
if hybrid_entities or hybrid_relations:
layers.append("hybrid")
# Mergear regex + hybrid evitando duplicados exactos (type_ref, name).
progress(0.85, "merge")
seen = set()
entities: list[dict] = []
for src in (regex_entities, hybrid_entities):
for e in src:
key = (e["type_ref"], e["name"])
if key in seen:
continue
seen.add(key)
entities.append(e)
if len(entities) >= max_entities:
break
if len(entities) >= max_entities:
break
# Reasignar tmp ids tras merge para que sean estables 0..N-1.
id_remap: dict[str, str] = {}
for i, e in enumerate(entities):
new_id = f"tmp_{i}"
id_remap[e["id"]] = new_id
e["id"] = new_id
relations: list[dict] = []
for r in hybrid_relations:
fi = id_remap.get(r["from_id"])
ti = id_remap.get(r["to_id"])
if fi is None or ti is None:
continue
relations.append({
"from_id": fi,
"to_id": ti,
"name": r.get("name") or "RELATED_TO",
"source": r.get("source") or "hybrid",
"confidence": float(r.get("confidence", 0.0)),
})
progress(1.0, "done")
print(json.dumps({
"entities": entities,
"relations": relations,
"stats": {
"layers": layers,
"n_entities": len(entities),
"n_relations": len(relations),
},
}, ensure_ascii=False))
return 0
if __name__ == "__main__":
sys.exit(main())