“张三上周跟‘字节’的所有沟通记录，发我一下。”
“王五手里金额超过 100 万的商机，最近有啥进展？”
“‘火山引擎’最近的续约意向有没有啥变化？”
在火山引擎客户经营平台 C360（Customer 360）上，每天都可以看到一线销售、销售管理者和运营在问这样的问题。听起来就是一次次普通检索，答案也就在系统里，可是关键词搜索却行不通，这是因为：
- “字节”可能指字节跳动，也可能指某家关联子公司；
- “上周”是相对时间，我们要先换算成日期；
- “续约意向有没有啥变化”的答案，我们要把对同一个客户最近的数条跟进记录读完，才能归纳总结好。
客户每一次拜访、每一轮沟通，都会被销售留在系统跟进记录里，作为管理者判断商机走向、发现流失风险时最依赖的数据。火山引擎 ContextSearch 团队基于 Agentic Search 改造了 C360 跟进记录的检索体验。经 200 道业务实际问答测试，Agentic Search 在整体正确率和可接受率上明显优于传统 RAG：正确率达到 84.0%，可接受率达到 90.0%，尤其在实体歧义处理场景、结构化精确查询和零答案问题上，正确率分别达到 91.7%、93.3% 和 100%。
本篇文章就聊三件事：关键词搜索及通用 RAG 在销售语料上为什么不好用？Agentic Search 靠怎样的链路把这些难题解决掉，以及用到业务里之后效果到底怎么样。
01关键词搜索在跟进记录上的三个难题
我们每个人的表达记录方式是不同的，所以跟进记录不可能按标准字段填写，销售用自己的话记录进展，记录里有客户简称、产品缩写、口语化描述和相对时间，还夹着合同附件、聊天截图、会议照片等等。
传统关键词搜索在这类语料上天然存在着三个躲不过去的难题：
- 语义鸿沟： 搜“字节跳动”，搜不到只写了“字节”的记录；“上周”“上个季度”这类相对时间系统不认识；产品缩写和销售口语基本等于零召回。
- 信息孤岛： 关键词只匹配正文中的文本内容，技术对接人、图片和备注里的关键信息可能在合同附件里，搜索做不到精准到位。
- 上下文断裂： 一个商机往往跟进数月、来回十几轮沟通，关键词搜索只能返回零散片段信息，没法围绕“这个客户”建立一个完整的过程。
如针对开头提到的每个查询，要实现正确的回复，都需要直面多个挑战：
02通用 RAG 解决了“相似”，却没解决“准确”
RAG 比关键词进一步：它能做语义匹配，让大模型基于召回的片段组织回答。对照三个难题看，它的能力边界很清楚——“字节跳动”能召回“字节”，语义鸿沟被部分填平；但信息孤岛和上下文断裂依旧，片段还是片段、问答仍是单轮。更关键的是，销售语料里有三道业务硬约束，通用 RAG 很难同时满足。
- 强业务语义。 问题里的客户、人员、团队、商机都是业务实体，这当中简称、同名、集团上下级公司充满了歧义。我们把“字节”向量化成一段语义，代替不了精确判断，我们还是无法确定“这个词在业务库里到底指哪家公司”的。
- 强权限约束。 销售只能看自己名下的客户，管理者能看整个团队，数据边界被业务系统定下了。权限必须在检索环节就生效，越权数据不召回；等答案生成后再删，来不及也不可靠。
- 强统计口径。 “华东区本季度跟进次数前十的客户”这类问题，答案要按明确口径从结构化数据里算出来，不能让大模型随便就从片段里估计出一个。
Agent 在这里有具体含义：系统自己承担“理解问题、拆解步骤、选择数据源、执行检索、组织证据”这条认知链路，承接了原来丢给用户的拼图工作。这也是 Agentic Search 和“接了大模型的搜索框”最根本的区别。
03Agentic Search 的完整链路
用户的每一次提问，实际上后台走的都是一条带反馈的完整链路：问题进来先做业务意图理解，再由实体对齐和提前建好的业务知识库共同支撑多路检索与过滤，结果按权限返回并附上证据，最后进入持续评测，反哺下一轮检索。
图一：Agentic Search 跟进记录场景业务链路
业务意图理解
Agent 拿到问题先分类：查明细，比如最近谁拜访过这个客户；做统计，比如某团队本季度跟进了多少次；做分析，比如这个客户有没有流失风险、对某款产品感不感兴趣。相对时间、客户简称、团队的口语叫法，都在这一步转成稳定的查询条件。
实体对齐
平台为客户、人员、团队、商机建了独立的实体库，支持拼音、中英文缩写、曾用名、关联公司等多维别名，配合专用的命名实体识别与纠错模型。“字节”指哪家公司由实体库判定，同名客户、集团上下级公司也能区分开。遇到名称或范围有歧义、答错代价高的情况，Agent 会先反问确认，而不是硬给一个可能误导人的答案。
业务知识库
业务知识库在用户提问前就建好了，检索时与实体对齐并行调用。跟进记录、客户信息、人员与组织关系统一接入，形成面向销售场景的业务知识库，查询口径与源业务系统保持一致。数据时效按两档保障：跟进记录分钟级增量同步，每天再做一次全量校准，上午写的记录下午就能被搜到。
多路检索与证据生成
关键词匹配、语义理解、结构化约束三条路并行。针对“上下文断裂”，检索结果按跟进人、客户 ID 聚合，把跨数月的十几轮沟通还原成一条完整时间线，而不是返回零散片段。明细类问题在时间线上定位原始记录，统计类问题走结构化约束按口径计算，分析类问题基于聚合后的记录归纳。每条结论都能溯源到具体数据（Grounding），用户可以顺着引用回去核对，大模型幻觉在这里没有多少空间。
权限过滤
所有查询沿用业务系统既有的权限体系（ACL），在检索阶段完成过滤，只返回当前用户有权查看的数据；检索链路完整留痕、可审计。人员转岗、客户移交后的权限变更随业务系统同步生效。
接入成本也压得很低：业务系统通过标准的 context-search-agent Skill 对接，只需配置 api_key 和 base_url 即可启用，不用改造现有数据链路。
至此，前面的提到传统搜索及通用 RAG 在销售语料上的难题和约束都有了明确的承接环节：
04业务效果
业务侧与 ContextSearch 团队共建了一套 200 道题的标准离线评测集，题目来自业务实际提问，覆盖精确查询、语义证据、零结果、歧义澄清等类型。经过测试，Agentic Search 的整体正确率高于传统检索增强方案。收益最明显的是这三类题：带精确条件的明细查询、客户与人员重名简称的歧义处理、按业务口径计算的统计问题。 开放式语义分析没有一味追求“模型说了算”，而是结论和证据结合，一起给出让销售自己判断答案是不是可行。
业务围绕着跟进记录检索和分析展开，9 月中旬统计，数据规模的覆盖已达到如下量级：
- 数百万级： 跟进记录样本，覆盖多年历史沟通信息
- 百万级： 客户覆盖，支持客户维度查询、统计与风险线索发现
- 千级： 销售与跟进人员，支持个人、团队、组织范围分析
ContextSearch 让 Agent 的每个判断都基于可追溯的跟进事实，把 C360 平台沉淀的数据真正变成了销售决策和行动的输入，当前已在风险预警、洞察提炼、能力传承等多类销售核心场景落地应用：
- 季度续约盘点： Agent 调出跟进记录，识别出“销售一直汇报、但是决策人三周没回复”的停滞客户，从而提前预警、推动干预，防流失。
- 新品反馈洞察： 新产品上线一个月后，Agent 归拢客户评价，提炼出“制造看权限、互联网看速度、销售不推”三个要点，从而优化点直接落到产品和话术。
- 新人交接： Agent 还原出对客的历史承诺和沟通结果，新销售一天内找到了“要逾期、没兑现、需重建联系”的三个问题，从而能快速上手，不重复问、不传错口径。
05经验与展望
跟进记录这一战，验证了 Agentic Search 的落地方式很适合强业务场景。对要建设同类系统的企业，有四点经验可以直接参考：
- 业务说人话，系统也得听懂人话： 不要要求用户按字段检索，简称、口语、相对时间由系统侧消化。
- 实体先对齐，再谈检索： 客户和人没对上，语义召回做得再好也是错上加错。
- 数字按口径算，结论给出处： 统计问题走结构化计算，分析问题附带原始记录，这样的答案才用的放心。
- 权限和评测与功能一起交付： 进真实业务系统，权限边界、审计留痕、离线回归这些一个都不能少。
销售跟进记录不是孤例，ContextSearch 高精度与复杂混合检索已经进入科研、法律、金融、电商等数十个行业，覆盖论文检索、类案检索、专利文书检索、保单检索、综合信息检索这些 “数据量大、行业术语多、权限分级、既要查事实又要做分析” 的场景。而搜索的评价标准也在随之改变：用户要的不再是“找到文档”，而是“问得出、搜得准”并能直接拿到“信得过”的答案。
了解更多 ContextSearch 相关信息，请点击这里。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
"Change all last week's communications with byte, send me."
"What's the latest progress on Wang Five's business offer of over a million dollars?"
“Have there been any changes in the intent of the recent renewal of Volcengine?”
On the Volcengine customer operating platform C360 (Customer 360), one-line sales managers and operations can be seen every day asking such questions. It sounds like a regular search, the answer is in the system, but a keyword search doesn't work because:
- “byte” may mean byte or a related subsidiary;
- Last week is relative time, and we have to start with convert;
- The answer to the question “Have there been any changes in the intention to renew the contract” is that we need to read the most recent follow-up notes on the same client in order to sum up.
Each visit by a client and every round of communication are kept in the system's follow-up records and are the data most relied upon by managers when they judge the course of business and identify the risk of loss. The Volcengine ContextSearch team has adapted the search experience of C360 follow-up records based on AgecSearch. After a practical question-and-answer test of 200 lines of operations, Agenic Search clearly outperformed the traditional RAG: 84.0 per cent correct and 90.0 per cent acceptable, particularly with regard to physical ambiguity, structured precision queries and zero answers, 91.7 per cent, 93.3 per cent and 100 per cent respectively.
There are three things in this article: "Keyword Search and Universal RAG" why not sell the language? What kind of chain does it take to solve these problems and how it works when it's used in the business?
#01 keyword search for three challenges on the follow-up record #
Each of us has a different expression, so it is not possible for a follow-up record to be completed in accordance with standard field, and the sale to record progress in its own words. It contains client abbreviations, product abbreviations, oral descriptions and relative time, as well as attachments to the contract, chats, meeting photographs, etc.
There are three natural difficulties with traditional keyword searches in such language:
- Semantic Gaps: Searching for “byte beats”, which does not include only “byte” records; relative time systems such as “last quarter” were not known last week; product abbreviations and sales language are almost equivalent to zero recall.
- Information island: Keywords match only text content in the body of the text, and key information from the technical interface, photos and notes may be in the contract attachments and cannot be accurately searched.
- context Breaks: A business opportunity often follows a few months, back and forth over a dozen rounds of communication. Keyword search can only return sporadic information and cannot create a complete process around “this customer”.
For each of the inquiries mentioned at the outset, the right response requires facing multiple challenges:
02 Generic RAG solved "similar" but not "accurate"
RAG goes further than the keyword: it can match semantics, allowing LLM to respond based on a recall segment. By contrast, its capacity boundaries are clear — “byte beats” can recall “byte”, and the semantic divide is partially closed; however, the information island and context break apart and the segment and question-and-answer sessions remain one-way. More crucially, there are three hard-bargaining lines of business in the distribution language, and it's hard to satisfy the generic RAG at the same time.
- Strong business semantics. The clients, people, teams and business opportunities at issue are business entities, and the acronyms, the same name, and the top and bottom companies of the group are all mixed. We can't substitute the semantics of “byte” to quantization, and we still have no way of determining which company it is in the business library.
- Strong permission constraints. Sales can only be made by clients in their own name, managers can see the entire team, and data boundaries are set by the business system. permission must be effective at the retrieval link, and ultra vires data will not be recalled; it will not be reliable until the answer has been generated.
- Strong statistical caliber. The answer to questions such as “Customs of the top 10 follow-up in East China this quarter” needs to be calculated from the structured data by a clear calibration, and LLM cannot simply estimate one from the clip.
Agent has a specific meaning here: the system itself carries the cognitive link of “understanding problems, dismantling steps, selecting data sources, executing retrieval, organizing evidence”, taking over the puzzle that was originally dumped on users. This is also the most fundamental difference between Agenic Search and "the search box with LLM".
Complete link to 03AgenticSearch
Each user ' s question, in fact, follows a complete chain of feedback: questions come in to understand business intentions, then multiple search and filtration are jointly supported by entity alignment and a pre-established business knowledge base, resulting in a return to permission with supporting evidence, and finally an ongoing assessment and a follow-up round of searches.
Figure I: Institutional Search Follow-up to the recording scene business link
Business intent understanding
Agent gets a sort of problem first: identify the details, for example, who recently visited the client; do statistics, for example, how many times a team has followed up this quarter; do analyses, for example, whether the client is at risk of being lost and not interested in a certain product. The relative timing, client abbreviations, and the group's accent are moving towards stable search conditions.
Entity Alignment
The Platform has built a stand-alone physical repository for clients, people, teams, business opportunities, supporting multi-dimensional aliases such as spelling, Chinese and English abbreviations, past names, associated companies, etc., in conjunction with a dedicated named entity to identify and correct errors. “byte” refers to companies that are determined by the physical bank and can be distinguished from the same customers, group superiors and subordinates. In cases where the name or scope is ambiguous and the cost of the wrong answer is high, Agent will ask for confirmation first, rather than hard enough to give a potentially misleading answer.
Business knowledge base
The business knowledge base is built before user questions are asked and is retrieved in parallel with the entity. Follow-up of records, client information, and integrated access to people-organization relationships, resulting in a business knowledge base oriented to the sales scene, with calibration and source operating systems aligned. Data validity is guaranteed in two tranches: follow-up of minutes of record increments, full calibration once a day, and morning records are searched in the afternoon.
Multi-road search and evidence generation
Keyword matching, semantic understanding, structured binding three parallel paths. For “context break-ups”, the results of the search are aggregated by follow-up, client ID, returning dozens of rounds of communication over several months to a complete time line rather than returning to a fragmentary segment. Questions of the breakdown are based on the original records in the time line, questions of the statistical category are structured by calibration, and questions of the analytical category are based on the consolidated records. Each conclusion can be traced back to specific data (Grounding), and users can cross-reference back and check, and there is little space for LLM hallucinations.
permission Filter
All queries follow the existing permission system (ACL) of the operating system, which is filtered at the retrieval stage and returns only the data that the current user is entitled to view; the search links are completely traceable and auditable. Changes to permission following the transfer of personnel and clients take effect simultaneously with the operational system.
Access costs are also very low: the business system is connected by standard contact-search-agent Skill, which can be activated with api key and base url, without re-engineering the existing data links.
So, the previous reference to traditional search and the common RAG has a clear link to the difficulties and constraints of marketing:
04 Operational effectiveness
In conjunction with the ContextSearch team, a standard offline assessment set of 200 topical issues was developed on the side of the operation, covering the types of precise queries, semantic evidence, zero results, ambiguity clarifications, etc. After testing, the overall correctness of Agenic Seach is higher than that of the traditional search enhancement programme. The most obvious are the three types of issues: detailed inquiries with precise conditions, vague treatment of clients and persons, and statistical questions based on business calibre. The open semantic analysis does not simply seek “models to figure out”, but rather a combination of conclusions and evidence, which, together, gives the sale the possibility of judging itself whether it is feasible.
Operations revolved around follow-up to record retrieval and analysis, and in mid-September the coverage of data scales reached the following levels:
- Millions: follow up on sample, covering years of historical communication.
- Million-degree: client coverage to support client dimension queries, statistics and risk leads
- Thousands: Sales and follow-up staff, supporting individuals, teams, organizational scope analysis
ContextSearch has allowed every judgment of Agent to be based on retroactive follow-up, turning the data from the C360 station into a real input into sales decisions and actions, which are currently being applied in a variety of core sales scenarios, such as risk warning, insulation, and capacity inheritance:
• Quarterly renewals: Agent pulls out follow-up records and identifies stagnating clients who have been reporting sales but who have not responded for three weeks, thereby providing early warning, facilitating intervention and preventing leakage.
- New product feedback: One month after the new product went online, Agent came together to evaluate the client, extracting the three main points of “manufacturing to see permission, Internet to see speed, sales without pushing”, thus optimizing the point to product and myth.
- New person transfer: Agent returns the historical promise and communication of the client, and within one day of the new sale, three questions were identified: “To be overdue, to be unattainable, to be reconnected”, so that he can move quickly, without repeating questions, without miscalculation.
05 Experiences and perspectives
Following up on this war, it was verified that the landing of Agenic Search was well suited to a strong business landscape. There are four direct points of reference for enterprises that are developing similar systems:
- The system also has to understand the language of business: do not require users to search it by field, abbreviations, words, relative time digested by the system.
- The entity aligns first, then the search: the customer and the person are wrong, the semantic recall is wrong.
- The figures are calibrated and the conclusions given: the statistical questions are calculated in a structured manner and the analysis of the questions is accompanied by the original records.
- permission and evaluation and function delivery: Into the real business system, permission border, audit trails, offline returns, none less.
Sales follow-up records are not unique, and ContextSearch's high-precision and complex combination of search has entered dozens of industries, such as scientific research, legal, financial, electrical and other, covering the search of papers, type search, patent document search, policy retrieval, comprehensive information retrieval, such as “the volume of data, the number of trade terms, permission rankings, both factual and analytical”. The evaluation criteria of the search have changed accordingly: instead of finding the document, the user wants the answer to "required, searched " and can obtain "trusted".
Please click here for more information about ContextSearch.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.