叙事导航者 · Narrative Planner

01 — THE PROBLEM

通用大模型在长程访谈中容易失焦 General-purpose LLMs lose focus in long-form interviews.

回忆录访谈是一项专业工作。当对话动辄数十轮时，我们观察到现有 LLM 普遍存在三类系统性缺陷： Memoir interviewing is a craft. Across long, multi-turn conversations we observe three recurring failure modes in off-the-shelf LLMs:

PAIN · 01

遗忘主线Losing the thread

缺乏全局规划。话题随机漂移，重要的人生阶段被遗漏，访谈结束时仍存在大片空白。 No global plan. Topics drift, key life stages get skipped, and entire decades remain blank when the session ends.

PAIN · 02

提问浅尝辄止Shallow probing

缺乏深度挖掘机制。停在表面事实，无法触及情感、细节与反思——而那才是回忆录的灵魂。 Stops at surface facts. Never digs into emotion, sensory detail, or reflection — the substance that makes a memoir worth reading.

PAIN · 03

逻辑跳跃Jarring transitions

话题切换生硬。从父亲跳到工作再跳到童年，老人感到困惑，对话失去连贯。 Topic switches feel abrupt. The interviewee gets disoriented, and the conversation loses narrative coherence.

02 — OUR APPROACH

把"规划"和"说话"分开 Decouple planning from speaking.

我们设计了一个独立的访谈决策规划模块（Planner），它不直接生成对话，而是输出结构化的战术指令——告诉对话 Agent 下一步该深挖什么、何时切换话题、用什么语气。Planner 的依据来自一张实时构建、不断生长的人生事件图谱。 We built a standalone interview decision planner that does not produce dialogue itself. Instead it emits structured tactical instructions — telling the conversational agent what to probe next, when to switch topics, and what tone to use. Its decisions are grounded in a live, ever-growing life-event graph.

"用科技留存岁月。" "Preserving lives through technology."

三阶段演化 · Three-Stage Evolution of the Life Graph Three-Stage Evolution of the Life Graph

STAGE 01

骨架期Skeleton

从极简生平出发，搭建"待填空的生命树"——通用人生阶段大纲。 Start from minimal biography. Build a blank life-tree — a universal stage-by-stage outline waiting to be filled.

STAGE 02

蔓延生长期Growth

边访谈边织网：捕捉高亮实体（人名、时间），挂载到骨架，预测潜在子任务。 Weave the web while interviewing: capture salient entities (people, times), attach to the skeleton, predict latent sub-tasks.

STAGE 03

主题升华期Synthesis

覆盖率 > 70% 时切换至总结模式，扫描全图寻找高频情绪，提炼出主题（如《漂泊中的安宁》）。 Once coverage exceeds 70%, switch to synthesis mode: scan the graph for emotional motifs and surface a theme (e.g., "Stillness Within Drifting").

03 — HOW IT WORKS

四层架构，五个 Agent 协同 Four layers. Five cooperating agents.

NAVIGATION · A

深度扩展Deep Dive

当核心事件槽位未满、情绪能量高时触发。三条子路径： Triggered when core slots are unfilled and emotional intensity is high. Three sub-paths:

A1由事及人Facts → People — 归纳法，提炼性格标签induction, distill character traits
A2由物及情Object → Emotion — 感官锚点触发深度记忆sensory anchors unlock deep memory
A3由时代及个体Era → Individual — 历史背景中的个人坐标locate the person within history

NAVIGATION · B

广度跳转Breadth Switch

当信息密度衰减、出现疲劳时触发。五种跳转策略，按体验质量排序： Triggered when information density drops or fatigue appears. Five strategies, ordered by interview quality:

B1情感共振 — 通过情绪钩子串联跨度大的事件Affective resonance — link distant events via shared emotion
B2社交网络 — 利用图谱中已识别但未展开的人物Social network — pivot through people already in the graph
B3地理跳转 — 沿物理空间（家、学校、工厂）延伸Spatial jump — follow places (home, school, factory)
B4时代快照 — 抛出标志性历史时刻定位Historical snapshot — anchor on a landmark moment
B5挂起任务 — 翻出此前被打断的话题，保障覆盖率Task resume — pick up an earlier interrupted thread

04 — KEY FEATURES

从规划到产出，端到端 End-to-end, planning to publication.

实时交互Real-Time

基于 WebSocket 的流式回应，规划与对话同步进行，无感等待。 WebSocket-streamed responses. Planning runs concurrently with dialogue — no waiting.

动态图谱Dynamic Graph

基于 Neo4j + GraphRAG 实时构建人生事件图谱，覆盖率可视化。 Live event graph on Neo4j + GraphRAG. Coverage and density rendered visually.

多模态Multimodal

支持照片上传、语音口音识别、多图叙事链条，触发感官记忆。 Photo upload, accent-aware speech, multi-image narrative chains — sensory cues that unlock memory.

叙事风格Narrative Voice

三种产出风格：个人风格、名家风格、传统回忆录风格。 Three output voices: personal, literary master, classic memoir.

无幻觉No Hallucination

所有产出严格基于访谈内容，每段叙述可追溯到原始对话。 Every line of output is grounded in actual interview content. Fully traceable to source dialogue.

评估体系Evaluation Suite

覆盖率、深度、连贯性的自动化指标，配 LLM-as-judge 仿真测试。 Automated coverage, depth, and coherence metrics, paired with LLM-as-judge simulations.

断点续传Resumable

老人可随时中断，下次开启自动温故知新，从挂起状态继续。 The interviewee can pause anytime. Next session resumes from the saved state — gracefully.

柔性引导Soft Guidance

话题切换使用桥接语，避免审讯感，让老人感到被倾听。 Topic switches use bridging language. No interrogation feel — the interviewee feels heard.

05 — DEMO

看看它的样子 See it in action.

06 — EVALUATION

对比基线，量化提升 Measured against the baseline.

我们构建了 LLM 驱动的"模拟受访者"环境，让 Planner 与无规划的纯对话 Agent 进行多轮访谈对照实验。 We built an LLM-driven simulated interviewee and ran controlled multi-turn experiments comparing the Planner against a vanilla conversational baseline.

话题覆盖率Coverage

+10.3% vs baseline

覆盖更完整的人生阶段，减少访谈空白。 More complete life-stage coverage, fewer narrative gaps.

对话深度Avg. Depth

+381.7% vs baseline

每个话题的平均挖掘深度显著提升。 Significantly deeper exploration per topic.

逻辑连贯性Coherence

+17.3% LLM judge

由 LLM 评分的话题切换流畅度。 Topic-switch fluency, scored by LLM as judge.

※ 完整实验报告与原始数据见技术报告 PDF。 ※ Full experimental report and raw data available in the technical paper.

叙事导航者 Narrative
Planner