diff --git a/.env.example b/.env.example
index b19a7fa..4ec4661 100644
--- a/.env.example
+++ b/.env.example
@@ -29,6 +29,12 @@ CSRF_TRUSTED_ORIGINS=http://localhost:5173,http://127.0.0.1:5173
 # OPENAI_TEXT_MODEL=
 # 别名：LLM_API_KEY、LLM_BASE_URL、LLM_MODEL
 
+# 报告/策略等「纯文本」任务的后端选择（默认经 crawler 副本内 AI_crawler 调 OpenAI 兼容网关；其它取值待扩展）
+# MA_LLM_TEXT_PROVIDER=crawler_openai_compatible
+
+# 独立策略稿 / 报告内「策略与机会」LLM 块采样温度（默认 0.1，低于 chat_completion_text 的 0.2，减轻同提示多轮漂移）。设 0 更稳、设 0.2 与旧默认接近。
+# MA_STRATEGY_LLM_TEMPERATURE=0.1
+
 # --- 可选：流水线侧 LLM 开关 ---
 # MA_SKIP_LLM_KEYWORD_SUGGEST=1
 # MA_ENABLE_LLM_COMMENT_SENTIMENT=1
diff --git a/backend/pipeline/llm/keyword_suggest.py b/backend/pipeline/llm/keyword_suggest.py
index ad5f58d..5dd697b 100644
--- a/backend/pipeline/llm/keyword_suggest.py
+++ b/backend/pipeline/llm/keyword_suggest.py
@@ -1,21 +1,17 @@
-"""在报告生成前：基于评价正文调用大模型，联想补充**关注词**与**使用场景触发组**（写入 effective_report_config）。"""
+"""在报告生成前：基于评价正文调用大模型，联想**短语候选**（写入 keyword_suggest_llm.json；不再合并进预设词表）。"""
 from __future__ import annotations
 
 import json
 import re
-import sys
-from pathlib import Path
 from typing import Any
 
-from django.conf import settings
+from .llm_client import call_llm
 
 MAX_CHUNK_CHARS = 24_000
 MAX_CHUNKS = 12
-# 场景联想单次送入模型的评价摘录上限（字符）；过大易顶上下文
-SCENARIO_CORPUS_MAX_CHARS = 18_000
 
 _CHUNK_SYSTEM = """你是电商评价挖掘助手。输入 JSON 含 keyword、excerpt_index、excerpts（一段用户评价正文合集）。
-任务：从 excerpts 中抽取值得纳入「关注词/卖点监测」的**中文短语**（2～12 字为主，可为词组）。
+任务：从 excerpts 中抽取**可人工选用的监测短语**（卖点、体验、规格等，**非**系统预设词表；2～12 字为主，可为词组）。
 
 硬性规则：
 - 仅输出一段 JSON：{"phrases": ["短语1", ...]}，短语共 6～20 条。
@@ -24,25 +20,6 @@ _CHUNK_SYSTEM = """你是电商评价挖掘助手。输入 JSON 含 keyword、ex
 - 不要输出 JSON 以外的文字。"""
 
 
-def _ensure_ai_crawler_path() -> None:
-    root = Path(settings.CRAWLER_JD_ROOT).resolve()
-    if not root.is_dir():
-        raise FileNotFoundError(f"爬虫副本目录不存在: {root}")
-    rs = str(root)
-    if rs not in sys.path:
-        sys.path.insert(0, rs)
-
-
-def _call_llm(system_prompt: str, user_prompt: str) -> str:
-    _ensure_ai_crawler_path()
-    import AI_crawler as ac  # noqa: WPS433
-
-    return ac.chat_completion_text(
-        system_prompt=system_prompt,
-        user_prompt=user_prompt,
-    )
-
-
 def _chunk_comment_texts(texts: list[str]) -> list[str]:
     """将全量评价划为若干段，控制单段字符量与最大段数。"""
     parts: list[str] = []
@@ -92,127 +69,6 @@ def _parse_phrases_object(raw: str) -> list[str]:
     return []
 
 
-_SCENARIO_SYSTEM = """你是电商用户研究助手。输入 JSON 含：
-- ``keyword``：监测词；
-- ``existing_scenarios``：数组，每项为 ``{"label": "展示名", "triggers": ["子串1", ...]}``。统计时若评价正文**包含任一 trigger 子串**，则计入该 label（与宿主系统规则一致）。
-- ``excerpts``：多条用户评价正文摘录（已截断拼接）。
-
-任务：在**不重复** ``existing_scenarios`` 中已有 ``label``（逐字比较，勿改写字）的前提下，从 excerpts 归纳 **4～12 条**新的「用途/场景」监测组，覆盖评论里**明显出现但未被现有组覆盖**的消费情境（如「下午茶」「露营」「宿舍」等，须确有文本依据）。
-
-硬性规则：
-- **仅输出**一段 JSON：``{"scenarios": [{"label": "展示名", "triggers": ["子串1", "子串2", ...]}, ...]}``；
-- 每条 ``label`` 2～16 字；每组 ``triggers`` 3～10 条，每条 trigger 为 **2～12 字中文**子串，用于**子串命中**计数；
-- 不要医疗功效、治愈、降血糖承诺；不要与 existing 的 label 同名或仅差空格；
-- 不要输出 JSON 以外的文字。"""
-
-
-def _sample_corpus_for_scenarios(texts: list[str], *, max_chars: int) -> str:
-    """取评价正文前部拼接至 max_chars，供单次场景联想。"""
-    parts: list[str] = []
-    n = 0
-    for t in texts:
-        s = (t or "").strip()
-        if not s:
-            continue
-        extra = len(s) + 1
-        if n + extra > max_chars:
-            remain = max_chars - n - 1
-            if remain > 40:
-                parts.append(s[:remain])
-            break
-        parts.append(s)
-        n += extra
-    return "\n".join(parts)
-
-
-def _parse_scenarios_object(raw: str) -> list[dict[str, Any]]:
-    t = (raw or "").strip()
-    t = re.sub(r"^```(?:json)?\s*", "", t, flags=re.IGNORECASE)
-    t = re.sub(r"\s*```$", "", t)
-    try:
-        obj = json.loads(t)
-    except json.JSONDecodeError:
-        obj = None
-    if not isinstance(obj, dict):
-        m = re.search(r"\{[\s\S]*\}", t)
-        if not m:
-            return []
-        try:
-            obj = json.loads(m.group(0))
-        except json.JSONDecodeError:
-            return []
-    arr = obj.get("scenarios")
-    if not isinstance(arr, list):
-        return []
-    out: list[dict[str, Any]] = []
-    for item in arr:
-        if not isinstance(item, dict):
-            continue
-        label = str(item.get("label") or "").strip()[:80]
-        tr_raw = item.get("triggers")
-        triggers: list[str] = []
-        if isinstance(tr_raw, list):
-            seen_t: set[str] = set()
-            for x in tr_raw[:24]:
-                s = str(x).strip()
-                if len(s) < 2 or len(s) > 24:
-                    continue
-                if s in seen_t:
-                    continue
-                seen_t.add(s)
-                triggers.append(s)
-        if label and len(triggers) >= 1:
-            out.append({"label": label, "triggers": triggers[:12]})
-    return out
-
-
-def suggest_scenario_groups_llm(
-    *,
-    keyword: str,
-    existing_groups: list[dict[str, Any]],
-    all_comment_texts: list[str],
-) -> dict[str, Any]:
-    """
-    单次调用模型，基于评价摘录扩展 ``comment_scenario_groups`` 形态的新组（label + triggers）。
-    """
-    if not all_comment_texts:
-        return {
-            "suggested_scenario_groups": [],
-            "scenario_rationale": "无评价正文可分析。",
-        }
-    existing_compact: list[dict[str, Any]] = []
-    for g in (existing_groups or [])[:36]:
-        if not isinstance(g, dict):
-            continue
-        lab = str(g.get("label") or "").strip()
-        tr = g.get("triggers")
-        ts: list[str] = []
-        if isinstance(tr, list):
-            for x in tr[:16]:
-                s = str(x).strip()
-                if s:
-                    ts.append(s[:48])
-        if lab and ts:
-            existing_compact.append({"label": lab[:80], "triggers": ts})
-    excerpts = _sample_corpus_for_scenarios(
-        all_comment_texts, max_chars=SCENARIO_CORPUS_MAX_CHARS
-    )
-    payload = {
-        "keyword": keyword,
-        "existing_scenarios": existing_compact,
-        "excerpts": excerpts,
-    }
-    raw = _call_llm(_SCENARIO_SYSTEM, json.dumps(payload, ensure_ascii=False))
-    scenarios = _parse_scenarios_object(raw)
-    return {
-        "suggested_scenario_groups": scenarios[:14],
-        "scenario_rationale": (
-            f"基于约 {len(excerpts)} 字评价摘录单次调用模型；"
-            f"在 {len(existing_compact)} 组既有场景之外补充 {len(scenarios[:14])} 组候选。"
-        ),
-    }
-
-
 def suggest_focus_keywords_from_all_comments(
     *,
     keyword: str,
@@ -245,7 +101,7 @@ def suggest_focus_keywords_from_all_comments(
             "excerpt_index": i + 1,
             "excerpts": ch,
         }
-        raw = _call_llm(_CHUNK_SYSTEM, json.dumps(payload, ensure_ascii=False))
+        raw = call_llm(_CHUNK_SYSTEM, json.dumps(payload, ensure_ascii=False))
         collected.extend(_parse_phrases_object(raw))
 
     seen: set[str] = set()
@@ -264,7 +120,7 @@ def suggest_focus_keywords_from_all_comments(
         "suggested_focus_keywords": out_kw,
         "rationale": (
             f"基于全量 {len(all_comment_texts)} 条评价文本，分 {len(chunks)} 段调用模型抽取短语并去重；"
-            f"已排除与当前关注词统计表完全相同的词。"
+            "报告主文不以子串词表统计为主指标，本结果仅供业务人工参考。"
         ),
         "chunks_processed": len(chunks),
         "total_comment_texts": len(all_comment_texts),
diff --git a/backend/pipeline/llm/llm_client.py b/backend/pipeline/llm/llm_client.py
index 4237f1a..88b18a4 100644
--- a/backend/pipeline/llm/llm_client.py
+++ b/backend/pipeline/llm/llm_client.py
@@ -1,47 +1,30 @@
-"""竞品报告 LLM 调用：路径注入与网关 ``chat_completion_text`` 封装。"""
+"""竞品报告 LLM 调用：经 `providers` 工厂选择后端，并统一对输出做去围栏等归一化。"""
 from __future__ import annotations
 
-import os
-import sys
-from pathlib import Path
-
-from django.conf import settings
+from .providers.factory import get_text_llm
+from .providers.output_normalize import strip_outer_markdown_fence
 
 
-def ensure_ai_crawler_path() -> None:
-    root = Path(settings.CRAWLER_JD_ROOT).resolve()
-    if not root.is_dir():
-        raise FileNotFoundError(f"爬虫副本目录不存在: {root}")
-    rs = str(root)
-    if rs not in sys.path:
-        sys.path.insert(0, rs)
-
-
-def call_llm(system_prompt: str, user_prompt: str) -> str:
-    ensure_ai_crawler_path()
-    import AI_crawler as ac  # noqa: WPS433
-
-    raw = ac.chat_completion_text(
-        system_prompt=system_prompt,
-        user_prompt=user_prompt,
+def call_llm(
+    system_prompt: str,
+    user_prompt: str,
+    *,
+    temperature: float | None = None,
+) -> str:
+    raw = get_text_llm().complete_text(
+        system_prompt,
+        user_prompt,
+        temperature=temperature,
     )
-    return ac.strip_outer_markdown_fence(raw)
+    return strip_outer_markdown_fence(raw)
 
 
 def estimate_chat_input_tokens(system_prompt: str, user_prompt: str) -> int:
-    """与 ``AI_crawler._estimate_chat_input_tokens`` 一致，用于在调用前预判上下文。"""
-    total_chars = len(system_prompt or "") + len(user_prompt or "")
-    return int(total_chars * 0.55) + 512
+    """与当前所选文本后端的预检一致；默认与 ``AI_crawler`` 的保守估算同口径。"""
+    return get_text_llm().estimate_input_tokens(system_prompt, user_prompt)
 
 
 def llm_context_window_size() -> int:
-    """与 ``AI_crawler.chat_completion_text`` 使用的上下文上限一致。"""
-    raw = (
-        os.environ.get("LLM_CONTEXT_WINDOW")
-        or os.environ.get("OPENAI_CONTEXT_WINDOW")
-        or "32768"
-    ).strip()
-    try:
-        return max(4096, int(raw))
-    except ValueError:
-        return 32768
+    """与当前所选后端的上下文上限一致；默认与 ``AI_crawler.chat_completion_text`` 使用的环境变量一致。"""
+    return get_text_llm().context_window_tokens()
+
diff --git a/backend/pipeline/llm/providers/__init__.py b/backend/pipeline/llm/providers/__init__.py
new file mode 100644
index 0000000..e8032da
--- /dev/null
+++ b/backend/pipeline/llm/providers/__init__.py
@@ -0,0 +1,11 @@
+"""文本大模型调用的协议、适配器与工厂（与具体提示词/业务生成逻辑解耦）。"""
+from __future__ import annotations
+
+from .factory import get_text_llm, reset_text_llm_client_for_tests
+from .protocol import TextLlmClient
+
+__all__ = [
+    "TextLlmClient",
+    "get_text_llm",
+    "reset_text_llm_client_for_tests",
+]
diff --git a/backend/pipeline/llm/providers/crawler_openai_compatible.py b/backend/pipeline/llm/providers/crawler_openai_compatible.py
new file mode 100644
index 0000000..ac47146
--- /dev/null
+++ b/backend/pipeline/llm/providers/crawler_openai_compatible.py
@@ -0,0 +1,64 @@
+"""
+经 `crawler_copy/jd_pc_search/AI_crawler.chat_completion_text` 访问 OpenAI 兼容网关（与配料识别等共用凭据与配置）。
+"""
+from __future__ import annotations
+
+import os
+import sys
+from pathlib import Path
+
+from django.conf import settings
+
+from .token_heuristics import estimate_crawler_style_input_tokens
+
+
+def ensure_ai_crawler_path() -> None:
+    root = Path(settings.CRAWLER_JD_ROOT).resolve()
+    if not root.is_dir():
+        raise FileNotFoundError(f"爬虫副本目录不存在: {root}")
+    rs = str(root)
+    if rs not in sys.path:
+        sys.path.insert(0, rs)
+
+
+def _llm_context_window_size_from_env() -> int:
+    raw = (
+        os.environ.get("LLM_CONTEXT_WINDOW")
+        or os.environ.get("OPENAI_CONTEXT_WINDOW")
+        or "32768"
+    ).strip()
+    try:
+        return max(4096, int(raw))
+    except ValueError:
+        return 32768
+
+
+class CrawlerOpenAiCompatibleTextLlm:
+    """
+    文本任务默认后端：复用 `AI_crawler` 的 `chat_completion_text` 与上下文预检逻辑。
+    更换为其它云厂商时，应新增独立适配器并在 `factory` 中注册，而非修改本类。
+    """
+
+    def complete_text(
+        self,
+        system_prompt: str,
+        user_prompt: str,
+        *,
+        temperature: float | None = None,
+    ) -> str:
+        ensure_ai_crawler_path()
+        import AI_crawler as ac  # noqa: WPS433
+
+        kwargs: dict[str, object] = {
+            "system_prompt": system_prompt,
+            "user_prompt": user_prompt,
+        }
+        if temperature is not None:
+            kwargs["temperature"] = float(temperature)
+        return ac.chat_completion_text(**kwargs)
+
+    def estimate_input_tokens(self, system_prompt: str, user_prompt: str) -> int:
+        return estimate_crawler_style_input_tokens(system_prompt, user_prompt)
+
+    def context_window_tokens(self) -> int:
+        return _llm_context_window_size_from_env()
diff --git a/backend/pipeline/llm/providers/factory.py b/backend/pipeline/llm/providers/factory.py
new file mode 100644
index 0000000..42a7119
--- /dev/null
+++ b/backend/pipeline/llm/providers/factory.py
@@ -0,0 +1,48 @@
+"""
+根据 `MA_LLM_TEXT_PROVIDER` 选择文本大模型实现；未设置时与历史行为一致（经 AI_crawler 的 OpenAI 兼容网关）。
+"""
+from __future__ import annotations
+
+import os
+
+from .crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm
+from .protocol import TextLlmClient
+
+# 模块级单例：避免重复构造；测试可用 `reset_text_llm_client_for_tests` 切换实现。
+_client: TextLlmClient | None = None
+
+# 与历史默认行为一致
+_DEFAULT_ID = "crawler_openai_compatible"
+_ENV_KEY = "MA_LLM_TEXT_PROVIDER"
+
+
+def _provider_id() -> str:
+    raw = (os.environ.get(_ENV_KEY) or _DEFAULT_ID).strip().lower()
+    return raw or _DEFAULT_ID
+
+
+def _build_client(pid: str) -> TextLlmClient:
+    if pid in (
+        "crawler_openai_compatible",
+        "crawler",
+        "default",
+        "openai_compatible",
+    ):
+        return CrawlerOpenAiCompatibleTextLlm()
+    raise ValueError(
+        f"不支持的 {_ENV_KEY}={pid!r}；"
+        f"当前仅实现 {_DEFAULT_ID}（经 AI_crawler 的 OpenAI 兼容 `chat/completions`）。"
+    )
+
+
+def get_text_llm() -> TextLlmClient:
+    global _client
+    if _client is None:
+        _client = _build_client(_provider_id())
+    return _client
+
+
+def reset_text_llm_client_for_tests() -> None:
+    """供 pytest/集成测试在修改环境变量后清空缓存的客户端。"""
+    global _client
+    _client = None
diff --git a/backend/pipeline/llm/providers/output_normalize.py b/backend/pipeline/llm/providers/output_normalize.py
new file mode 100644
index 0000000..6b10532
--- /dev/null
+++ b/backend/pipeline/llm/providers/output_normalize.py
@@ -0,0 +1,15 @@
+"""对模型原始输出做与业务无关的轻量归一化（不修改提示词）。"""
+from __future__ import annotations
+
+
+def strip_outer_markdown_fence(text: str) -> str:
+    """若模型用 ``` / ```markdown 包裹全文，去掉最外层围栏。与 `AI_crawler.strip_outer_markdown_fence` 行为一致。"""
+    t = (text or "").strip()
+    if not t.startswith("```"):
+        return t
+    lines = t.split("\n")
+    if lines and lines[0].strip().startswith("```"):
+        lines = lines[1:]
+    while lines and lines[-1].strip() == "```":
+        lines = lines[:-1]
+    return "\n".join(lines).strip()
diff --git a/backend/pipeline/llm/providers/protocol.py b/backend/pipeline/llm/providers/protocol.py
new file mode 100644
index 0000000..3060908
--- /dev/null
+++ b/backend/pipeline/llm/providers/protocol.py
@@ -0,0 +1,22 @@
+"""文本 LLM 客户端协议：业务侧只依赖本接口，具体网关由适配器 + 工厂选择。"""
+from __future__ import annotations
+
+from typing import Protocol, runtime_checkable
+
+
+@runtime_checkable
+class TextLlmClient(Protocol):
+    def complete_text(
+        self,
+        system_prompt: str,
+        user_prompt: str,
+        *,
+        temperature: float | None = None,
+    ) -> str:
+        """一次 system + user 的纯文本补全，返回助理正文（无通用后处理，由 `llm_client.call_llm` 统一去围栏等）。"""
+
+    def estimate_input_tokens(self, system_prompt: str, user_prompt: str) -> int:
+        """与当次后端的 `max_tokens` 预检/截断策略一致的输入侧 token 保守估算。"""
+
+    def context_window_tokens(self) -> int:
+        """当前配置下的上下文 token 上限（与预检、策略模块档位一致）。"""
diff --git a/backend/pipeline/llm/providers/token_heuristics.py b/backend/pipeline/llm/providers/token_heuristics.py
new file mode 100644
index 0000000..302fe85
--- /dev/null
+++ b/backend/pipeline/llm/providers/token_heuristics.py
@@ -0,0 +1,10 @@
+"""
+与 `crawler_copy/.../AI_crawler` 中 `_estimate_chat_input_tokens` 同口径的保守估算，
+供预检、策略档位与 OpenAI 兼容适配器共用（无 tiktoken 时避免 max_tokens 400）。
+"""
+from __future__ import annotations
+
+
+def estimate_crawler_style_input_tokens(system_prompt: str, user_prompt: str) -> int:
+    total_chars = len(system_prompt or "") + len(user_prompt or "")
+    return int(total_chars * 0.55) + 512