dev/issues/README.md

@@ -45,7 +45,7 @@
 | [0037](completed/0037-ioc-regex-extractor.md) | IoC regex extractor (IP, email, dominio, hash, wallet, CVE, MAC) | completado | alta | feature | — |
 | [0038](completed/0038-gliner-entity-extractor.md) | GLiNER entity extractor (zero-shot NER multilingue) | completado | alta | feature | 0039, 0040 |
 | [0039](completed/0039-glirel-relation-extractor.md) | GLiREL relation extractor (zero-shot triplets) | completado | media | feature | 0040 |
-| [0040](0040-hybrid-extraction-pipeline.md) | Pipeline hibrido extraccion grafos (regex + GLiNER + GLiREL + LLM fallback) | pendiente | media | feature | — |
+| [0040](completed/0040-hybrid-extraction-pipeline.md) | Pipeline hibrido extraccion grafos (regex + GLiNER + GLiREL + LLM fallback) | completado | media | feature | — |
 | [0041](completed/0041-cpp-app-best-practices.md) | C++ app shell estandarizado (PATTERNS.md + AppConfig extendido) | completado | alta | feature | 0043 |
 | [0042](completed/0042-cpp-layout-storage-public.md) | C++ layout_storage publico (extraer de shaders_lab) | completado | alta | feature | 0043 |
 | [0043](completed/0043-cpp-apps-standardize-shell.md) | Estandarizar shell de las 4 apps C++ | completado | alta | refactor | 0046 |

python/functions/pipelines/extract_graph_hybrid.md

@@ -0,0 +1,192 @@
+---
+name: extract_graph_hybrid
+kind: pipeline
+lang: py
+domain: pipelines
+version: "1.0.0"
+purity: impure
+signature: "def extract_graph_hybrid(chunks: list[str], entity_schema: list[dict], relation_types: list[str], gliner_model: Any, glirel_model: Any, llm_chat_json: Callable[[list[dict]], dict] | None = None, ioc_types: list[str] | None = None, confidence_threshold: float = 0.6, languages: str = 'Respond in Spanish.', min_entities_per_chunk: int = 2) -> tuple[list[EntityCandidate], list[RelationCandidate]]"
+description: "Pipeline hibrido en cascada que combina extract_iocs (regex, coste 0), GLiNER (zero-shot NER, coste bajo), GLiREL (zero-shot RE) y un LLM fallback opcional para chunks complejos o de baja confianza. Devuelve listas concatenadas listas para deduplicate_entities/deduplicate_relations."
+tags: [pipeline, extraction, entities, relations, gliner, glirel, ioc, regex, llm, nlp, datascience, cybersecurity, hybrid]
+uses_functions:
+  - extract_iocs_py_cybersecurity
+  - extract_entities_gliner_py_datascience
+  - extract_relations_glirel_py_datascience
+  - extract_entities_llm_py_datascience
+  - extract_relations_llm_py_datascience
+uses_types:
+  - entity_candidate_py_datascience
+  - relation_candidate_py_datascience
+returns:
+  - entity_candidate_py_datascience
+  - relation_candidate_py_datascience
+returns_optional: false
+error_type: "error_go_core"
+imports:
+  - typing.Any
+  - typing.Callable
+  - warnings
+params:
+  - name: chunks
+    desc: "Lista de fragmentos de texto ya cortados (p.ej. via split_text_into_chunks)."
+  - name: entity_schema
+    desc: "Schema para GLiNER y LLM. Lista de dicts con type_ref, label y opcional metadata_fields."
+  - name: relation_types
+    desc: "Tipos de relacion permitidos para GLiREL/LLM (ej: ['operates','owns','communicates_with'])."
+  - name: gliner_model
+    desc: "Instancia GLiNER cargada con gliner_load_model. Inyectada por el caller."
+  - name: glirel_model
+    desc: "Instancia GLiREL cargada con glirel_load_model. Inyectada por el caller."
+  - name: llm_chat_json
+    desc: "Cliente LLM inyectado (sin acoplamiento al proveedor). Si None, no hay fallback LLM."
+  - name: ioc_types
+    desc: "Subset de tipos para extract_iocs (email, ip_address, domain, file_hash, ...). None = todos."
+  - name: confidence_threshold
+    desc: "Por debajo de este umbral, GLiNER se considera de baja confianza y se invoca el LLM."
+  - name: languages
+    desc: "Instruccion de idioma passthrough al LLM (ej: 'Respond in Spanish.')."
+  - name: min_entities_per_chunk
+    desc: "Si un chunk arroja menos entidades que esto, se invoca el LLM como fallback (default 2)."
+output: "Tupla (entities, relations) con candidatas concatenadas (sin deduplicar). El caller debe pasar por deduplicate_entities y deduplicate_relations."
+tested: true
+tests:
+  - "corpus OSINT con IoCs y entidades semanticas devuelve mezcla regex+GLiNER"
+  - "chunks vacios o con solo whitespace se saltan"
+  - "entity_schema vacio lanza ValueError"
+  - "chunks no-lista lanza ValueError"
+  - "GLiNER produciendo pocas entidades dispara fallback LLM si llm_chat_json esta presente"
+  - "sin llm_chat_json no se invoca ningun fallback LLM"
+  - "GLiREL sin relaciones dispara fallback LLM relations"
+  - "ioc_types acota el set de extractores regex"
+  - "errores de extractores se capturan con warnings y no abortan el pipeline"
+test_file_path: "python/functions/pipelines/tests/test_extract_graph_hybrid.py"
+file_path: "python/functions/pipelines/extract_graph_hybrid.py"
+---
+
+## Ejemplo
+
+```python
+from python.functions.pipelines.extract_graph_hybrid import extract_graph_hybrid
+from python.functions.datascience.gliner_load_model import gliner_load_model
+from python.functions.datascience.glirel_load_model import glirel_load_model
+from python.functions.datascience.deduplicate_entities import deduplicate_entities
+from python.functions.datascience.deduplicate_relations import deduplicate_relations
+
+gliner = gliner_load_model("urchade/gliner_multi-v2.1", device="auto")
+glirel = glirel_load_model("jackboyla/glirel-large-v0", device="auto")
+
+entity_schema = [
+    {"type_ref": "osint_person_go_cybersecurity",       "label": "Person"},
+    {"type_ref": "osint_organization_go_cybersecurity", "label": "Organization"},
+    {"type_ref": "osint_location_go_cybersecurity",     "label": "Location"},
+]
+relation_types = ["operates", "owns", "communicates_with", "employed_by"]
+
+chunks = [
+    "Alice Johnson works at OpenAI in San Francisco. Contact: alice@openai.com.",
+    "The C2 server lives at 192.168.0.1 and resolves to evil-corp.com.",
+]
+
+# Sin LLM (coste cero, solo regex + GLiNER + GLiREL)
+entities, relations = extract_graph_hybrid(
+    chunks=chunks,
+    entity_schema=entity_schema,
+    relation_types=relation_types,
+    gliner_model=gliner,
+    glirel_model=glirel,
+    llm_chat_json=None,
+)
+
+# Con LLM fallback solo en chunks complejos
+def llm_chat_json(messages):
+    # llamar a OpenAI/Anthropic/Ollama y devolver el JSON ya parseado
+    ...
+
+entities, relations = extract_graph_hybrid(
+    chunks=chunks,
+    entity_schema=entity_schema,
+    relation_types=relation_types,
+    gliner_model=gliner,
+    glirel_model=glirel,
+    llm_chat_json=llm_chat_json,
+    confidence_threshold=0.6,
+    min_entities_per_chunk=2,
+)
+
+# Deduplicar antes de persistir
+dedup = deduplicate_entities(entities, name_threshold=0.85)
+final_relations = deduplicate_relations(relations, dedup.name_to_id)
+```
+
+## Algoritmo
+
+Por cada chunk:
+
+1. **Regex (capa tecnica)** — `extract_iocs(chunk, ioc_types)` devuelve dicts
+   `{value, start, end, type}` que se mapean a `EntityCandidate` con
+   `type_ref` propio (`ioc_email`, `ioc_ip_address`, `ioc_domain`, ...) y
+   `confidence=1.0`. Los offsets se anotan en `attributes['start'/'end']`
+   para que GLiREL pueda mapearlos a tokens sin fallback `text.find`.
+2. **GLiNER (capa semantica)** — `extract_entities_gliner` con el schema y
+   el `confidence_threshold` como filtro de score.
+3. **Merge** — IoCs + GLiNER deduplicados por `(name, type_ref)`. NO se
+   colapsa fuzzy aqui; eso lo hace el caller.
+4. **LLM fallback (opcional)** — si el chunk tiene menos de
+   `min_entities_per_chunk` entidades **o** `mean(gliner_confidence) <
+   confidence_threshold` **y** `llm_chat_json is not None`, se invoca
+   `extract_entities_llm` y se mezcla.
+5. **GLiREL (relaciones zero-shot)** — solo si hay >=2 entidades.
+6. **LLM fallback de relaciones (opcional)** — si GLiREL no devolvio nada
+   con >=2 entidades **y** hay `llm_chat_json`, se invoca
+   `extract_relations_llm` para ese chunk.
+
+`source_chunk_indices` y `source_chunk_index` se rellenan para que
+`deduplicate_relations` pueda reconstruir el grafo origen→destino.
+
+## Por que cascada y no all-LLM
+
+| Capa | Coste por 100 KB | Latencia | Calidad |
+|------|------------------|----------|---------|
+| `extract_iocs` (regex) | 0 | <50 ms | Precision 100% en IoCs tecnicos |
+| GLiNER (`gliner_multi-v2.1`) | 0 (modelo local, GPU/CPU) | ~1-3 s/chunk en CPU, <0.5 s en GPU | F1 0.7-0.85 en NER zero-shot |
+| GLiREL (`glirel-large-v0`) | 0 (modelo local) | ~2-4 s/chunk en CPU | F1 0.5-0.75 en RE zero-shot |
+| LLM (GPT-4 / Claude Sonnet) | $0.5-3 por 100 KB | 5-15 s/chunk | F1 0.85-0.95 |
+
+El pipeline hibrido reserva el LLM (caro y lento) para los chunks que
+GLiNER/GLiREL no resuelven con suficiente confianza. En corpus OSINT
+tipicos el LLM se invoca en <20% de los chunks → coste total 5-10x menor
+que un pipeline 100% LLM con perdida de calidad <5 puntos F1.
+
+## Solapamiento IoC ↔ GLiNER
+
+GLiNER puede detectar `apple.com` como `Organization` mientras que regex
+lo detecta como `domain`. **Decision intencional**: ambos se conservan
+con `type_ref` distinto (`osint_organization_go_cybersecurity` vs
+`ioc_domain`). `deduplicate_entities(..., same_type_only=True)` no las
+mezcla. El caller decide si quiere unificar (por ejemplo, anotando una
+relacion `domain_of` entre las dos).
+
+## Recomendaciones operativas
+
+- **Batch size**: ~100-200 chunks de 500-1000 caracteres por llamada al
+  pipeline. Mas chunks → mas paralelismo aprovechable; menos chunks →
+  menos overhead de carga del modelo.
+- **Latencia esperada (CPU)**: ~3-5 s/chunk sin LLM, +5-15 s/chunk si
+  cae al LLM fallback.
+- **Latencia esperada (GPU)**: ~0.5-1 s/chunk sin LLM.
+- **Cuando bajar `confidence_threshold`**: en corpus con jerga muy
+  especifica donde GLiNER no aprendio bien — pero esto incrementa el
+  coste si hay LLM (mas chunks caen al fallback).
+- **Cuando subir `min_entities_per_chunk`**: si quieres forzar fallback
+  LLM en chunks "ricos" para asegurar cobertura completa.
+
+## Notas
+
+- La deduplicacion fuzzy (Levenshtein + Union-Find) la hace
+  `deduplicate_entities` — NO replicar aqui.
+- Los errores de cualquier extractor en cualquier chunk se capturan con
+  `warnings.warn` y NO abortan el pipeline (robustez sobre completitud).
+- Las funciones LLM aceptan `language_instruction`; aqui se pasa como
+  `languages` (default `"Respond in Spanish."`).
+- Pensar en una app `apps/osint_extractor/` que use este pipeline + sigma
+  viz como demo. Fuera de scope de este issue.

python/functions/pipelines/extract_graph_hybrid.py

@@ -0,0 +1,260 @@
+"""Pipeline hibrido: extract_iocs + GLiNER + GLiREL + LLM fallback."""
+
+from __future__ import annotations
+
+import os
+import sys
+import warnings
+from typing import Any, Callable
+
+_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", ".."))
+if _ROOT not in sys.path:
+    sys.path.insert(0, _ROOT)
+
+from python.functions.cybersecurity.extract_iocs import extract_iocs
+from python.functions.datascience.extract_entities_gliner import extract_entities_gliner
+from python.functions.datascience.extract_relations_glirel import extract_relations_glirel
+from python.functions.datascience.extract_entities_llm import extract_entities_llm
+from python.functions.datascience.extract_relations_llm import extract_relations_llm
+from python.types.datascience.entity_candidate import EntityCandidate
+from python.types.datascience.relation_candidate import RelationCandidate
+
+
+_IOC_TYPE_REF = {
+    "email": "ioc_email",
+    "ip_address": "ioc_ip_address",
+    "domain": "ioc_domain",
+    "file_hash": "ioc_file_hash",
+    "crypto_wallet": "ioc_crypto_wallet",
+    "cve_id": "ioc_cve_id",
+    "mac_address": "ioc_mac_address",
+    "phone_number": "ioc_phone_number",
+}
+
+_IOC_LABEL = {
+    "email": "Email",
+    "ip_address": "IPAddress",
+    "domain": "Domain",
+    "file_hash": "FileHash",
+    "crypto_wallet": "CryptoWallet",
+    "cve_id": "CVE",
+    "mac_address": "MACAddress",
+    "phone_number": "PhoneNumber",
+}
+
+
+def _ioc_dict_to_candidate(ioc: dict, chunk_index: int) -> EntityCandidate:
+    """Convierte un dict de extract_iocs a EntityCandidate.
+
+    Anota offsets en `attributes['start'/'end']` para que extract_relations_glirel
+    pueda mapearlos a tokens sin fallback `text.find`.
+    """
+    ioc_type = ioc.get("type", "")
+    return EntityCandidate(
+        name=ioc.get("value", ""),
+        type_ref=_IOC_TYPE_REF.get(ioc_type, f"ioc_{ioc_type}"),
+        type_label=_IOC_LABEL.get(ioc_type, ioc_type),
+        attributes={"start": ioc.get("start", -1), "end": ioc.get("end", -1)},
+        confidence=1.0,
+        source_chunk_indices=[chunk_index],
+    )
+
+
+def _mean_confidence(entities: list[EntityCandidate]) -> float:
+    if not entities:
+        return 0.0
+    return sum(e.confidence for e in entities) / len(entities)
+
+
+def _merge_entities_dedup_by_name_type(
+    base: list[EntityCandidate],
+    extra: list[EntityCandidate],
+) -> list[EntityCandidate]:
+    """Anade `extra` a `base` evitando duplicados exactos (name + type_ref).
+
+    No usa fuzzy: la deduplicacion final la hace el caller con
+    `deduplicate_entities`. Aqui solo evita el caso trivial de meter dos veces
+    la misma cadena con el mismo type_ref dentro del mismo chunk.
+    """
+    seen = {(e.name, e.type_ref) for e in base}
+    out = list(base)
+    for e in extra:
+        key = (e.name, e.type_ref)
+        if key in seen:
+            continue
+        seen.add(key)
+        out.append(e)
+    return out
+
+
+def extract_graph_hybrid(
+    chunks: list[str],
+    entity_schema: list[dict],
+    relation_types: list[str],
+    gliner_model: Any,
+    glirel_model: Any,
+    llm_chat_json: Callable[[list[dict]], dict] | None = None,
+    ioc_types: list[str] | None = None,
+    confidence_threshold: float = 0.6,
+    languages: str = "Respond in Spanish.",
+    min_entities_per_chunk: int = 2,
+) -> tuple[list[EntityCandidate], list[RelationCandidate]]:
+    """Extrae triplets `(entidad, relacion, entidad)` combinando regex + GLiNER + GLiREL + LLM fallback.
+
+    Cascada por chunk:
+      1. `extract_iocs(chunk, ioc_types)` → entidades tecnicas (precision 100%, coste 0).
+      2. `extract_entities_gliner(chunk, entity_schema, gliner_model)` → semanticas zero-shot.
+      3. Si entidades < `min_entities_per_chunk` o `mean(confidence) < confidence_threshold`
+         **y** hay `llm_chat_json` → `extract_entities_llm` para rellenar gaps.
+      4. `extract_relations_glirel(chunk, entidades_chunk, relation_types, glirel_model)`.
+      5. Si no salieron relaciones con >=2 entidades **y** hay `llm_chat_json` →
+         `extract_relations_llm` para esos chunks.
+
+    Args:
+        chunks: Lista de fragmentos de texto a procesar (ya tokenizados/cortados).
+        entity_schema: Schema para GLiNER y LLM. Lista de dicts con
+            `type_ref`, `label` y opcional `metadata_fields`.
+        relation_types: Tipos de relacion permitidos para GLiREL/LLM.
+        gliner_model: Instancia GLiNER cargada con `gliner_load_model`.
+        glirel_model: Instancia GLiREL cargada con `glirel_load_model`.
+        llm_chat_json: Funcion inyectada que recibe messages OpenAI-style y
+            retorna dict JSON. Si es None, no se invoca fallback LLM (ahorro maximo).
+        ioc_types: Subset de tipos para `extract_iocs`. None = todos.
+        confidence_threshold: Bajo este umbral se invoca el LLM como fallback.
+        languages: Instruccion de idioma para el LLM (passthrough a las funciones LLM).
+        min_entities_per_chunk: Si un chunk tiene menos entidades que esto,
+            se considera "complejo" y se llama al LLM.
+
+    Returns:
+        Tupla `(entities, relations)` con todas las candidatas concatenadas.
+        El caller debe pasar por `deduplicate_entities` y `deduplicate_relations`
+        antes de persistir. Cada `EntityCandidate` lleva
+        `source_chunk_indices=[i]` y cada `RelationCandidate`
+        lleva `source_chunk_index=i`.
+
+    Raises:
+        ValueError: Si `entity_schema` esta vacio o `chunks` no es lista.
+    """
+    if not isinstance(chunks, list):
+        raise ValueError("chunks debe ser una lista")
+    if not entity_schema:
+        raise ValueError("entity_schema no puede estar vacio")
+
+    all_entities: list[EntityCandidate] = []
+    all_relations: list[RelationCandidate] = []
+
+    for i, chunk in enumerate(chunks):
+        if not chunk or not chunk.strip():
+            continue
+
+        # ── Capa 1: regex IoCs ──────────────────────────────────────────────
+        try:
+            ioc_dicts = extract_iocs(chunk, ioc_types)
+        except Exception as exc:
+            warnings.warn(
+                f"extract_graph_hybrid: extract_iocs fallo en chunk {i}: {exc}",
+                stacklevel=2,
+            )
+            ioc_dicts = []
+
+        ioc_entities = [_ioc_dict_to_candidate(d, i) for d in ioc_dicts]
+
+        # ── Capa 2: GLiNER ──────────────────────────────────────────────────
+        try:
+            gliner_entities = extract_entities_gliner(
+                text=chunk,
+                entity_schema=entity_schema,
+                model=gliner_model,
+                threshold=confidence_threshold,
+            )
+        except Exception as exc:
+            warnings.warn(
+                f"extract_graph_hybrid: extract_entities_gliner fallo en chunk {i}: {exc}",
+                stacklevel=2,
+            )
+            gliner_entities = []
+
+        for ent in gliner_entities:
+            if i not in ent.source_chunk_indices:
+                ent.source_chunk_indices.append(i)
+
+        chunk_entities = _merge_entities_dedup_by_name_type(ioc_entities, gliner_entities)
+
+        # ── Capa 3: LLM entity fallback (opcional) ──────────────────────────
+        needs_entity_llm = (
+            len(chunk_entities) < min_entities_per_chunk
+            or _mean_confidence(gliner_entities) < confidence_threshold
+        )
+        if needs_entity_llm and llm_chat_json is not None:
+            try:
+                llm_entities = extract_entities_llm(
+                    text=chunk,
+                    entity_schema=entity_schema,
+                    llm_chat_json=llm_chat_json,
+                    language_instruction=languages,
+                )
+            except Exception as exc:
+                warnings.warn(
+                    f"extract_graph_hybrid: extract_entities_llm fallo en chunk {i}: {exc}",
+                    stacklevel=2,
+                )
+                llm_entities = []
+
+            for ent in llm_entities:
+                if i not in ent.source_chunk_indices:
+                    ent.source_chunk_indices.append(i)
+
+            chunk_entities = _merge_entities_dedup_by_name_type(chunk_entities, llm_entities)
+
+        all_entities.extend(chunk_entities)
+
+        # ── Capa 4: GLiREL ──────────────────────────────────────────────────
+        if len(chunk_entities) >= 2:
+            try:
+                glirel_relations = extract_relations_glirel(
+                    text=chunk,
+                    entities=chunk_entities,
+                    relation_types=relation_types,
+                    model=glirel_model,
+                    threshold=confidence_threshold,
+                )
+            except Exception as exc:
+                warnings.warn(
+                    f"extract_graph_hybrid: extract_relations_glirel fallo en chunk {i}: {exc}",
+                    stacklevel=2,
+                )
+                glirel_relations = []
+        else:
+            glirel_relations = []
+
+        for rel in glirel_relations:
+            rel.source_chunk_index = i
+
+        # ── Capa 5: LLM relation fallback (opcional) ────────────────────────
+        if (
+            llm_chat_json is not None
+            and len(chunk_entities) >= 2
+            and not glirel_relations
+        ):
+            try:
+                llm_relations = extract_relations_llm(
+                    text=chunk,
+                    entities=chunk_entities,
+                    relation_types=relation_types,
+                    llm_chat_json=llm_chat_json,
+                    language_instruction=languages,
+                )
+            except Exception as exc:
+                warnings.warn(
+                    f"extract_graph_hybrid: extract_relations_llm fallo en chunk {i}: {exc}",
+                    stacklevel=2,
+                )
+                llm_relations = []
+
+            for rel in llm_relations:
+                rel.source_chunk_index = i
+            glirel_relations.extend(llm_relations)
+
+        all_relations.extend(glirel_relations)
+
+    return all_entities, all_relations

python/functions/pipelines/tests/test_extract_graph_hybrid.py

@@ -0,0 +1,293 @@
+"""Tests de integracion para extract_graph_hybrid.
+
+Stubs duck-typed para gliner/glirel/LLM permiten ejercitar la cascada
+sin descargar modelos pesados.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Any
+
+import pytest
+
+_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", ".."))
+if _ROOT not in sys.path:
+    sys.path.insert(0, _ROOT)
+
+from python.functions.pipelines.extract_graph_hybrid import extract_graph_hybrid
+from python.types.datascience.entity_candidate import EntityCandidate
+from python.types.datascience.relation_candidate import RelationCandidate
+
+
+# ── Stubs ──────────────────────────────────────────────────────────────────────
+
+
+@dataclass
+class StubGliner:
+    """Stub de GLiNER. `responses` se va consumiendo por chunk en orden."""
+
+    responses: list[list[dict]] = field(default_factory=list)
+    calls: int = 0
+
+    def predict_entities(self, text, labels, threshold, flat_ner):
+        idx = self.calls
+        self.calls += 1
+        if idx < len(self.responses):
+            return self.responses[idx]
+        return []
+
+
+@dataclass
+class StubGlirel:
+    """Stub de GLiREL. Mismo patron que StubGliner."""
+
+    responses: list[list[dict]] = field(default_factory=list)
+    calls: int = 0
+
+    def predict_relations(self, tokens, labels, threshold, ner, top_k=1):
+        idx = self.calls
+        self.calls += 1
+        if idx < len(self.responses):
+            return self.responses[idx]
+        return []
+
+
+@dataclass
+class StubLLM:
+    """LLM stub: enruta por contenido del system prompt."""
+
+    entity_responses: list[dict] = field(default_factory=list)
+    relation_responses: list[dict] = field(default_factory=list)
+    entity_calls: int = 0
+    relation_calls: int = 0
+
+    def __call__(self, messages: list[dict]) -> dict:
+        system = messages[0]["content"] if messages else ""
+        if "relation extraction expert" in system.lower():
+            idx = self.relation_calls
+            self.relation_calls += 1
+            if idx < len(self.relation_responses):
+                return self.relation_responses[idx]
+            return {"relations": []}
+        idx = self.entity_calls
+        self.entity_calls += 1
+        if idx < len(self.entity_responses):
+            return self.entity_responses[idx]
+        return {"entities": []}
+
+
+SCHEMA = [
+    {"type_ref": "osint_person_go_cybersecurity",       "label": "Person"},
+    {"type_ref": "osint_organization_go_cybersecurity", "label": "Organization"},
+    {"type_ref": "osint_location_go_cybersecurity",     "label": "Location"},
+]
+RELATION_TYPES = ["operates", "owns", "communicates_with", "employed_by", "related_to"]
+
+
+# ── Tests ──────────────────────────────────────────────────────────────────────
+
+
+def test_corpus_osint_devuelve_mezcla_regex_gliner():
+    """Corpus OSINT con IoCs y entidades semanticas devuelve mezcla regex+GLiNER."""
+    chunks = [
+        "Alice Johnson works at OpenAI. Contact: alice@openai.com",
+    ]
+    gliner = StubGliner(responses=[
+        [
+            {"start": 0, "end": 13, "text": "Alice Johnson", "label": "Person", "score": 0.92},
+            {"start": 23, "end": 29, "text": "OpenAI", "label": "Organization", "score": 0.88},
+        ],
+    ])
+    glirel = StubGlirel(responses=[[]])
+
+    entities, relations = extract_graph_hybrid(
+        chunks=chunks,
+        entity_schema=SCHEMA,
+        relation_types=RELATION_TYPES,
+        gliner_model=gliner,
+        glirel_model=glirel,
+        llm_chat_json=None,
+    )
+
+    types = {e.type_ref for e in entities}
+    # Regex IoC: email
+    assert any(e.type_ref == "ioc_email" and e.name == "alice@openai.com" for e in entities)
+    # GLiNER: persona y organizacion
+    assert "osint_person_go_cybersecurity" in types
+    assert "osint_organization_go_cybersecurity" in types
+    # source_chunk_indices marcado
+    assert all(0 in e.source_chunk_indices for e in entities)
+    assert relations == []
+
+
+def test_chunks_vacios_se_saltan():
+    """Chunks vacios o solo whitespace se saltan sin invocar modelos."""
+    gliner = StubGliner(responses=[])
+    glirel = StubGlirel(responses=[])
+    entities, relations = extract_graph_hybrid(
+        chunks=["", "   ", "\n\t"],
+        entity_schema=SCHEMA,
+        relation_types=RELATION_TYPES,
+        gliner_model=gliner,
+        glirel_model=glirel,
+    )
+    assert entities == []
+    assert relations == []
+    assert gliner.calls == 0
+    assert glirel.calls == 0
+
+
+def test_entity_schema_vacio_lanza_value_error():
+    """entity_schema vacio lanza ValueError."""
+    with pytest.raises(ValueError):
+        extract_graph_hybrid(
+            chunks=["text"],
+            entity_schema=[],
+            relation_types=RELATION_TYPES,
+            gliner_model=StubGliner(),
+            glirel_model=StubGlirel(),
+        )
+
+
+def test_chunks_no_lista_lanza_value_error():
+    """chunks no-lista lanza ValueError."""
+    with pytest.raises(ValueError):
+        extract_graph_hybrid(
+            chunks="no soy lista",  # type: ignore[arg-type]
+            entity_schema=SCHEMA,
+            relation_types=RELATION_TYPES,
+            gliner_model=StubGliner(),
+            glirel_model=StubGlirel(),
+        )
+
+
+def test_gliner_pocas_entidades_dispara_fallback_llm():
+    """GLiNER produciendo pocas entidades dispara fallback LLM."""
+    chunks = ["Texto complejo sin patrones obvios."]
+    gliner = StubGliner(responses=[[]])  # GLiNER no encuentra nada
+    glirel = StubGlirel(responses=[[]])
+    llm = StubLLM(entity_responses=[
+        {"entities": [
+            {"name": "Acme Corp", "type_ref": "osint_organization_go_cybersecurity",
+             "attributes": {}, "confidence": 0.95},
+            {"name": "Bob",       "type_ref": "osint_person_go_cybersecurity",
+             "attributes": {}, "confidence": 0.9},
+        ]},
+    ])
+
+    entities, _ = extract_graph_hybrid(
+        chunks=chunks,
+        entity_schema=SCHEMA,
+        relation_types=RELATION_TYPES,
+        gliner_model=gliner,
+        glirel_model=glirel,
+        llm_chat_json=llm,
+        min_entities_per_chunk=2,
+    )
+
+    names = {e.name for e in entities}
+    assert "Acme Corp" in names
+    assert "Bob" in names
+    assert llm.entity_calls == 1
+
+
+def test_sin_llm_no_se_invoca_fallback():
+    """Sin llm_chat_json no se invoca ningun fallback LLM aunque GLiNER no encuentre nada."""
+    gliner = StubGliner(responses=[[]])
+    glirel = StubGlirel(responses=[[]])
+    entities, relations = extract_graph_hybrid(
+        chunks=["chunk dificil"],
+        entity_schema=SCHEMA,
+        relation_types=RELATION_TYPES,
+        gliner_model=gliner,
+        glirel_model=glirel,
+        llm_chat_json=None,
+    )
+    # Nada de LLM, solo lo que diera regex (en este chunk: nada)
+    assert entities == []
+    assert relations == []
+
+
+def test_glirel_sin_relaciones_dispara_fallback_llm_relations():
+    """GLiREL sin relaciones dispara fallback LLM relations."""
+    chunks = ["Alice Johnson trabaja para OpenAI."]
+    gliner = StubGliner(responses=[
+        [
+            {"start": 0, "end": 13, "text": "Alice Johnson", "label": "Person", "score": 0.95},
+            {"start": 26, "end": 32, "text": "OpenAI", "label": "Organization", "score": 0.9},
+        ],
+    ])
+    glirel = StubGlirel(responses=[[]])  # GLiREL no encuentra relaciones
+    llm = StubLLM(relation_responses=[
+        {"relations": [
+            {"from_name": "Alice Johnson", "to_name": "OpenAI",
+             "relation_type": "employed_by", "description": "...", "confidence": 0.9},
+        ]},
+    ])
+
+    _, relations = extract_graph_hybrid(
+        chunks=chunks,
+        entity_schema=SCHEMA,
+        relation_types=RELATION_TYPES,
+        gliner_model=gliner,
+        glirel_model=glirel,
+        llm_chat_json=llm,
+        confidence_threshold=0.5,
+        min_entities_per_chunk=2,
+    )
+
+    assert len(relations) == 1
+    assert relations[0].from_name == "Alice Johnson"
+    assert relations[0].to_name == "OpenAI"
+    assert relations[0].relation_type == "employed_by"
+    assert relations[0].source_chunk_index == 0
+    assert llm.relation_calls == 1
+
+
+def test_ioc_types_acota_extractores():
+    """ioc_types acota el set de extractores regex."""
+    chunks = ["Email: x@y.com, IP: 192.168.0.1, MD5: 5d41402abc4b2a76b9719d911017c592."]
+    gliner = StubGliner(responses=[[]])
+    glirel = StubGlirel(responses=[[]])
+    entities, _ = extract_graph_hybrid(
+        chunks=chunks,
+        entity_schema=SCHEMA,
+        relation_types=RELATION_TYPES,
+        gliner_model=gliner,
+        glirel_model=glirel,
+        llm_chat_json=None,
+        ioc_types=["email"],  # solo emails
+    )
+    types = {e.type_ref for e in entities}
+    assert "ioc_email" in types
+    assert "ioc_ip_address" not in types
+    assert "ioc_file_hash" not in types
+
+
+def test_errores_se_capturan_con_warning():
+    """Errores de extractores se capturan con warnings y no abortan el pipeline."""
+
+    class BoomGliner:
+        def predict_entities(self, *a, **k):
+            raise RuntimeError("boom")
+
+    class BoomGlirel:
+        def predict_relations(self, *a, **k):
+            raise RuntimeError("boom")
+
+    chunks = ["Email: contact@example.com"]
+    with pytest.warns(UserWarning):
+        entities, relations = extract_graph_hybrid(
+            chunks=chunks,
+            entity_schema=SCHEMA,
+            relation_types=RELATION_TYPES,
+            gliner_model=BoomGliner(),
+            glirel_model=BoomGlirel(),
+            llm_chat_json=None,
+        )
+    # Aun asi extract_iocs deberia haber sacado el email
+    assert any(e.type_ref == "ioc_email" for e in entities)
+    assert relations == []