diff --git a/docs/site/.vitepress/config.mts b/docs/site/.vitepress/config.mts index eaebef0..f19f7d3 100644 --- a/docs/site/.vitepress/config.mts +++ b/docs/site/.vitepress/config.mts @@ -87,6 +87,13 @@ export default defineConfig({ } ], '/ecosystem/': [ + { + text: 'WikiLoop', + items: [ + { text: 'WikiLoop in Industry', link: '/ecosystem/wikiloop-industries' }, + { text: 'Industry Use Cases', link: '/ecosystem/use-cases' }, + ] + }, { text: 'Systems', items: [ @@ -166,6 +173,13 @@ export default defineConfig({ } ], '/zh-CN/ecosystem/': [ + { + text: 'WikiLoop', + items: [ + { text: 'WikiLoop 行业应用', link: '/zh-CN/ecosystem/wikiloop-industries' }, + { text: '行业使用场景', link: '/zh-CN/ecosystem/use-cases' }, + ] + }, { text: '系统列表', items: [ diff --git a/docs/site/ecosystem/use-cases.md b/docs/site/ecosystem/use-cases.md new file mode 100644 index 0000000..f80d03d --- /dev/null +++ b/docs/site/ecosystem/use-cases.md @@ -0,0 +1,213 @@ +# Industry Use Cases + +How RAG and LLM Wiki knowledge bases are used across industries — what gets stored, who uses it, and which approach fits best. + +> **Quick guide:** +> - **RAG** — best for large, frequently updated document corpora requiring precise citation +> - **LLM Wiki** — best for stable, structured knowledge: concepts, SOPs, terminology, decision records +> - **Combined** — Wiki provides stable structure; RAG handles dynamic retrieval depth + +--- + +## Law Firms + +**What goes in the KB:** Case law and precedents, contract template libraries, M&A due diligence files, regulatory updates (SEC, GDPR, etc.), internal SOPs and matter management guides. + +| Use Case | Who | How | +|---|---|---| +| Legal research | Associates, paralegals | Natural language query returns cited case references | +| Contract review | Senior lawyers | Upload contract, system flags risk clauses vs. historical templates | +| M&A due diligence | Deal teams | Scan thousands of documents, extract key liabilities and risks | +| Compliance monitoring | Compliance officers | Auto-index regulatory updates, push relevant change alerts | +| Client self-service portal | Clients | Common legal questions answered from internal KB | + +**Approach:** RAG-primary (huge document volume, frequent regulation updates, citations required). LLM Wiki for legal concept definitions, internal terminology, standard process guides. Tools: Harvey AI, Thomson Reuters CoCounsel. + +--- + +## Accounting & Finance Firms + +**What goes in the KB:** GAAP/IFRS/tax code with interpretive guidance, IRS publications, audit standards (PCAOB/ISA), client financial statements and audit workpapers, SEC/FASB regulatory updates, standard engagement letter templates. + +| Use Case | Who | How | +|---|---|---| +| Tax research | Tax managers, accountants | Natural language query returns tax code citations with footnotes | +| Audit support | Audit teams | RAG retrieves audit standards, drafts audit memos | +| Financial statement analysis | Analysts, advisors | Upload 10-K/10-Q, LLM summarizes key risks and anomalies | +| Regulatory change tracking | Compliance | Auto-ingest FASB/PCAOB updates, push impact alerts | +| Client advisory chatbot | Wealth management clients | Portfolio-aware Q&A from personalized KB | + +**Approach:** RAG-primary (tax/standards volume is large, versions change). LLM Wiki for core GAAP concepts, internal operation manuals, common Q&A. Tools: Microsoft Copilot for Finance, Workiva AI, Bloomberg Tax AI. + +--- + +## Manufacturing Plants + +**What goes in the KB:** Equipment operation manuals, maintenance service bulletins, fault history and repair logs, SOPs, quality inspection checklists, OSHA/ISO safety standards, supplier specs and BOMs. + +| Use Case | Who | How | +|---|---|---| +| Equipment troubleshooting | Technicians on the floor | Describe fault in natural language, RAG returns manual steps | +| Quality control | QC staff | Query historical defect patterns, retrieve inspection SOPs | +| Safety compliance | Safety managers | Real-time retrieval of OSHA/ISO clauses for on-site issues | +| New employee onboarding | New workers | Conversational learning of plant procedures | +| Supplier & parts lookup | Procurement | Query lead times, part specs, and substitution options | + +**Approach:** RAG-primary for PDF-heavy equipment manuals. LLM Wiki for general plant operating norms, terminology dictionaries, onboarding paths. Combined: Wiki as structured backbone, RAG for specific manual details. + +--- + +## Software Development Teams + +**What goes in the KB:** Technical docs, API references, Architecture Decision Records (ADRs), Confluence/Notion pages, Jira ticket history, bug reports and resolutions, code comments and READMEs, key Slack/Teams discussions. + +| Use Case | Who | How | +|---|---|---| +| Developer doc Q&A | Engineers | Natural language query for API usage, architecture conventions | +| IT helpdesk automation | IT support | Tier 1/2 tickets auto-resolved from internal KB | +| New engineer onboarding | New hires | Ask about system design and code structure, get instant answers | +| Code review assist | Reviewers | AI retrieves similar historical PRs and decisions | +| Incident response | SRE/DevOps | Quickly retrieve runbooks, historical incident handling records | + +**Approach:** LLM Wiki-primary for architecture design, concepts, specs (ideal WikiLoop use case). RAG for unstructured docs (Confluence, Jira, Slack archives). Combined: WikiLoop holds stable knowledge; RAG handles latest Issues/PRs. Tools: GitHub Copilot Enterprise, Glean, Guru. + +--- + +## Customer Service Centers + +**What goes in the KB:** Product/service FAQs, refund/return policies, service agreements, historical tickets and resolved cases, agent scripts and escalation SOPs, CRM customer history. + +| Use Case | Who | How | +|---|---|---| +| Real-time agent assist | Support reps | RAG surfaces suggested answers during live calls | +| Self-service chatbot | End customers | Common questions answered by AI, reducing ticket volume | +| Ticket classification & routing | System | AI understands query content, routes to correct expert team | +| New agent training | New hires | Simulated Q&A to learn scripts and product knowledge | +| First-contact resolution | Quality team | Analyze knowledge gaps, continuously update KB | + +**Approach:** RAG-primary (policies update frequently with promotions and new products). LLM Wiki for stable standard scripts, return policies, escalation flows. Tools: Zendesk AI, Salesforce Einstein, Amazon Q. + +--- + +## Healthcare & Hospitals + +**What goes in the KB:** EHR summaries and patient history, clinical guidelines, drug interaction databases, ICD-10/CPT coding standards, hospital SOPs, nursing procedures, medical literature (PubMed). + +| Use Case | Who | How | +|---|---|---| +| Clinical decision support | Physicians | Query drug contraindications and treatment guidelines | +| Patient handoff summaries | Nurses, residents | AI auto-summarizes EHR for shift handoffs | +| Medical coding assist | Coders, billing | RAG matches historical cases to correct ICD/CPT codes | +| Patient FAQ bot | Patients | Self-service for appointments, fees, medication instructions | +| Radiology/pathology assist | Specialists | AI retrieves similar historical reports to assist draft | +| Drug information lookup | Pharmacists, nurses | Real-time formulary and drug database retrieval | + +**Approach:** RAG-primary (vast medical literature, fast-updating guidelines, evidence citations required). LLM Wiki for internal SOPs, nursing protocols, common care pathways. **Note:** HIPAA compliance requires private/on-premise deployment. Tools: Epic AI, Microsoft Azure Health Bot, Nuance DAX. + +--- + +## Financial Services & Investment Banking + +**What goes in the KB:** Regulatory docs (Basel III, MiFID II, Dodd-Frank), research reports, earnings call transcripts, trade history, risk model docs, KYC/AML compliance procedures, internal investment memos. + +| Use Case | Who | How | +|---|---|---| +| Investment research synthesis | Research analysts | RAG integrates multi-source reports, generates insight summaries | +| Compliance Q&A | Compliance / risk officers | Query latest regulatory text, assess business compliance | +| Client advisor support | Wealth advisors | Real-time product info and policy retrieval during calls | +| Anti-fraud knowledge base | Risk teams | Cross-team sharing of emerging fraud pattern knowledge | +| M&A due diligence | Investment bankers | Rapidly scan target company financial and legal documents | +| Trader compliance assist | Traders | Pre-trade compliance rule check to avoid violations | + +**Approach:** RAG-primary (high document volume, frequent regulatory updates). LLM Wiki for internal product knowledge, standardized compliance workflows. **Note:** Highly sensitive data requires air-gapped or private cloud RAG. + +--- + +## Education & Training + +**What goes in the KB:** Course content, teaching materials, enrollment policies, scholarship FAQs, exam syllabi, historical Q&A, faculty research outputs, accreditation compliance docs. + +| Use Case | Who | How | +|---|---|---| +| Student Q&A bot | Students | Self-service for course schedules, enrollment rules, scholarships | +| Personalized learning assistant | Students | AI tutoring based on course KB | +| Staff knowledge base | Faculty, admin | Query departmental policies, forms, compliance requirements | +| Corporate training retrieval | Employees | Search training materials, key points, historical case studies | +| Thesis research assist | Graduate students | Semantic search of related literature and existing findings | + +**Approach:** Combined — course structure and learning paths as Wiki; specific exercises and research literature via RAG. LLM Wiki-primary for standardized K12 content and fixed corporate training. Tools: Khanmigo, Microsoft Copilot for Education. + +--- + +## Government & Public Agencies + +**What goes in the KB:** Laws and regulations, policy documents, government service FAQs (tax, social security, visas), cross-department approval guides, historical policy interpretations, internal compliance regulations. + +| Use Case | Who | How | +|---|---|---| +| Citizen service portal | Public | 24×7 self-service for tax, social security, permit procedures | +| Policy document retrieval | Civil servants | Natural language search of policy text to support decisions | +| Cross-department knowledge sharing | All departments | Unified KB breaks information silos | +| Compliance review | Auditors | Retrieve regulations for side-by-side compliance audits | +| Parliamentary Q&A prep | Officials | Quickly retrieve historical documents to prepare responses | + +**Approach:** Combined — stable policies as structured Wiki for citizen understanding; full regulatory text indexed via RAG. **Note:** Data sovereignty requires on-premise deployment. Tools: Palantir AI Platform, Microsoft Azure Government. + +--- + +## E-commerce & Retail + +**What goes in the KB:** Product details, specifications, usage instructions, return/exchange policies, shipping FAQ, historical tickets, inventory and supplier data, promotion rules and membership benefits. + +| Use Case | Who | How | +|---|---|---| +| Smart product Q&A | Customers, agents | Natural language query for specs, stock, shipping time | +| Return/exchange automation | Customers | AI self-service for return requests, reducing manual handling | +| Personalized recommendations | Shoppers | KB combined with user profiles for precise product matching | +| Supply chain Q&A | Procurement, ops | Query inventory, supplier lead times, restocking suggestions | +| Agent assist | Support reps | Real-time policy answers during calls, reducing handle time | + +**Approach:** RAG-primary (massive SKU count, frequent product changes). LLM Wiki for stable return policies, membership tier rules, general FAQs. Tools: Salesforce Einstein, Zendesk AI, Shopify Sidekick. + +--- + +## More Industries + +| Industry | KB Contents | Key Use Cases | Approach | +|---|---|---|---| +| **Telecom** | Network fault guides, plan comparisons, churn intervention playbooks | Technician troubleshooting, customer plan recommendations, churn intervention scripts | RAG for faults, Wiki for scripts | +| **Insurance** | Policy terms, claims SOPs, underwriting standards, regulatory rules | Claim handler queries, underwriting assist, compliance Q&A | RAG-primary, Wiki for stable processes | +| **Life Sciences / Pharma** | Clinical trial data, FDA/EMA regulatory files, drug development docs | Researcher literature search, regulatory submission drafting | RAG + Wiki combined | +| **Real Estate** | Property listings, market reports, contract templates, local regulations | Agent Q&A, market analysis, contract review | RAG-primary | +| **Media & Publishing** | Archive articles, editorial guidelines, style guides, rights databases | Journalist research, editorial consistency checks, rights clearance | RAG + Wiki combined | + +--- + +## WikiLoop Specifically + +WikiLoop's local-first, MCP-native, FTS-based design fits a particular profile in each industry: + +| Industry | WikiLoop Fit | Specific Use | +|---|---|---| +| **Law firms** | High | Compile case research notes, decision records, client matter summaries into auditable Wiki | +| **Accounting firms** | High | Structured GAAP concept library, internal engagement methodology Wiki | +| **Manufacturing** | Medium | Plant SOPs, equipment glossary, onboarding knowledge base — stable content only | +| **Software dev** | Very High | ADR library, architecture knowledge base, team onboarding Wiki — WikiLoop's sweet spot | +| **Customer service** | Medium | Standard script Wiki, stable policy reference — combine with RAG for dynamic product data | +| **Healthcare** | High | Internal clinical SOPs, nursing pathway Wiki — private deployment satisfies HIPAA | +| **Financial services** | High | Internal compliance process Wiki, product knowledge base — air-gapped deployment | +| **Education** | High | Course structure Wiki, institutional knowledge base, research synthesis | +| **Government** | High | Policy interpretation Wiki, citizen FAQ base — sovereign deployment requirement met | +| **Research teams** | Very High | Literature synthesis, concept graphs, decision records — core WikiLoop use case | + +--- + +## Selection Guide + +| Factor | Choose RAG | Choose LLM Wiki | Choose Both | +|---|---|---|---| +| Document volume | Tens of thousands+ | Hundreds of pages | Mixed | +| Update frequency | Daily / weekly | Monthly / quarterly | Mixed cadence | +| Citation required | Yes (legal, medical) | Generally no | Sensitive domains | +| Content structure | Low (PDF, email) | High (process, concept) | Mixed formats | +| Query type | Open-ended retrieval | Fixed-pattern Q&A | Both types | diff --git a/docs/site/ecosystem/wikiloop-industries.md b/docs/site/ecosystem/wikiloop-industries.md new file mode 100644 index 0000000..dabffe4 --- /dev/null +++ b/docs/site/ecosystem/wikiloop-industries.md @@ -0,0 +1,186 @@ +# WikiLoop in Industry + +WikiLoop is a local-first knowledge search engine built for AI agents. Its architecture — a single binary, SQLite FTS5, plain Markdown, MCP protocol — makes it uniquely suited for industries with strict privacy, compliance, and air-gap requirements. + +## Why WikiLoop Fits High-Privacy Industries + +Most RAG systems require external infrastructure: cloud embedding APIs, vector databases, managed SaaS services. WikiLoop eliminates every one of these dependencies. + +| Requirement | How WikiLoop Meets It | +|---|---| +| **No cloud data transmission** | All data stays local. Zero external API calls for search or indexing. | +| **No embedding model required** | Pure SQLite FTS5 + BM25. No vector model, no embedding API, no GPU. | +| **Zero external infrastructure** | Single binary + one SQLite file. No Redis, Kafka, Postgres, or vector DB. | +| **Air-gap compatible** | stdio MCP mode runs as a local subprocess. Works with zero network access. | +| **Auditable knowledge** | All wiki pages are plain Markdown in git. Every change is a diff. | +| **No vendor lock-in** | OKF v0.1 compatible. KB is a directory of files — portable to any system. | +| **Local LLM support** | Distillation works with Ollama or any local model. Never sends documents to OpenAI. | +| **BYOK (Bring Your Own Key)** | If cloud LLM is used for distillation, only the API key is needed — no SaaS subscription. | + +--- + +## Law Firms + +**Privacy concern:** Attorney-client privilege. Client documents cannot leave the firm's infrastructure. + +**How WikiLoop is used:** + +- Compile case research notes, legal memos, and precedent analysis into structured wiki pages +- Build a firm-wide ADR-style decision library: past deal structures, litigation strategies, settlement patterns +- Each matter gets a source-note; concepts (legal theories, regulatory frameworks) synthesized into concept pages +- `wikiloop lint` validates citation integrity — every claim traceable to a source document +- Git history provides full audit trail for privilege review + +**Deployment:** Single binary on a firm-managed server or laptop. No internet required after initial setup. LLM distillation can use a local model (Ollama + Llama 3) or an on-premise API. + +--- + +## Accounting & Finance Firms + +**Privacy concern:** Client financial data under NDA, regulatory restrictions on data handling (SOX, GDPR). + +**How WikiLoop is used:** + +- Structured GAAP/IFRS concept library: each standard becomes a wiki page with `related_to` links to relevant interpretations +- Internal engagement methodology Wiki — how the firm approaches specific audit types, documented as decision pages +- Compile client-specific knowledge (anonymized) for cross-engagement learning without exposing raw client data +- Knowledge gap analysis (`wikiloop synthesize --gaps`) surfaces under-documented areas before peak season + +**Deployment:** On-premise or private cloud. Distillation with local model keeps client data entirely within firm boundaries. + +--- + +## Healthcare & Hospitals + +**Privacy concern:** HIPAA (US), GDPR (EU), local health data laws. Patient data cannot touch public cloud. + +**How WikiLoop is used:** + +- Internal clinical SOP Wiki: nursing protocols, care pathways, post-operative procedures — structured and searchable +- Drug formulary knowledge base: hospital-specific formulary compiled from standard references, kept up to date by the pharmacy team +- Incident and near-miss knowledge base: de-identified case summaries distilled into source-notes, concept pages for root cause patterns +- New resident onboarding: institution-specific procedures and protocols in a searchable Wiki, reducing reliance on senior staff +- Research team knowledge base: literature synthesis, trial design decisions, IRB-approved protocol documentation + +**Deployment:** Air-gapped hospital intranet. `wikiloop stdio` runs as a subprocess of the clinical AI tool. Zero network egress. + +--- + +## Financial Services & Investment Banking + +**Privacy concern:** Material non-public information (MNPI), trading confidentiality, regulatory requirements (MiFID II, SEC Rule 10b-5). + +**How WikiLoop is used:** + +- Internal compliance process Wiki: structured documentation of approved trading practices, escalation procedures, pre-clearance workflows +- Deal knowledge base: for each completed transaction, a source-note captures the deal structure, key decisions, and rationale — builds institutional memory across deal teams +- Regulatory interpretation library: compliance team maintains a Wiki of how the firm interprets specific regulatory rules, with `contradicts` links flagging conflicting interpretations +- Risk model documentation: quantitative analysts document model assumptions, limitations, and validation findings as decision pages + +**Deployment:** Air-gapped trading floor infrastructure or private cloud with no public internet path. Distillation uses on-premise LLM. + +--- + +## Manufacturing & Industrial + +**Privacy concern:** Trade secrets, proprietary manufacturing processes, IP in equipment design. + +**How WikiLoop is used:** + +- Plant SOP Wiki: standard operating procedures for each production line, searchable by process step, equipment type, or product family +- Equipment troubleshooting knowledge base: distilled from maintenance logs, service bulletins, and technician field notes — agents query it during repairs +- Quality defect pattern library: historical defect data compiled into source-notes; concept pages synthesize recurring root causes +- New employee onboarding: structured knowledge base reduces dependency on experienced workers for process knowledge transfer +- Supplier and materials knowledge: spec sheets, substitution history, and vendor reliability notes compiled as wiki pages + +**Deployment:** Factory floor OT network, often physically isolated from the internet. WikiLoop runs on a local server; technicians query via MCP through their AI coding assistant or custom chat interface. + +--- + +## Government & Defense + +**Privacy concern:** National security, classified information, data sovereignty requirements. + +**How WikiLoop is used:** + +- Policy interpretation Wiki: structured documentation of how specific regulations are applied in practice — searchable by policy area +- Citizen service knowledge base: FAQ and procedure guides for government service delivery, maintained by subject-matter experts +- Cross-agency knowledge sharing: common operational procedures compiled into a shared Wiki, reducing duplication across departments +- Procurement and compliance knowledge: past procurement decisions documented as decision pages, with rationale and alternative options considered + +**Deployment:** Sovereign cloud or on-premise with no external connectivity. `wikiloop serve` behind an internal network boundary. MCP over stdio for embedded agent use. OKF-compatible KB format enables knowledge transfer across secure systems. + +--- + +## Energy & Utilities (Nuclear, Oil & Gas) + +**Privacy concern:** Critical infrastructure protection, proprietary operational data, physical air-gap requirements. + +**How WikiLoop is used:** + +- Operations and maintenance knowledge base: equipment manuals, incident reports, and procedure updates distilled into structured wiki pages +- Safety and regulatory compliance Wiki: plant-specific interpretations of NRC/HSE regulations, emergency procedures, safety case documentation +- Knowledge retention for aging workforce: experienced engineers' tacit knowledge captured as source-notes and concept pages before retirement +- Incident learning library: post-incident reports distilled and cross-linked — `related_to` edges surface similar past events during future incident response + +**Deployment:** Physically air-gapped operational technology (OT) network. WikiLoop runs with zero external dependencies. Distillation batch-processed offline; index serves agent queries in real time. + +--- + +## Research Institutions & Think Tanks + +**Privacy concern:** Pre-publication research confidentiality, grant compliance, institutional IP. + +**How WikiLoop is used:** + +- Literature synthesis: researchers drop papers into `raw/`, WikiLoop distills source-notes and synthesizes concept and comparison pages across the corpus +- Research decision records: methodology choices, hypothesis revisions, data source evaluations documented as decision pages +- Cross-project knowledge sharing: common methodologies and findings compiled into a shared institutional Wiki +- Knowledge gap analysis: `wikiloop synthesize --gaps` identifies under-explored areas before grant proposal writing + +**WikiLoop's specific advantage here:** This is the closest to WikiLoop's original design intent. Agents search the knowledge base iteratively, follow `related` links across papers, and synthesize their own conclusions — exactly how a researcher works. + +**Deployment:** Researcher's own machine or institutional server. No cloud required. Local Ollama model for distillation keeps pre-publication data entirely offline. + +--- + +## Private Deployment Architecture + +For all industries above, a typical WikiLoop deployment looks like: + +``` +[Local KB directory] + raw/ ← drop source documents here + wiki/ ← distilled pages (Markdown + git) + index/ ← SQLite FTS index (auto-managed) + +[WikiLoop binary] + wikiloop serve ← HTTP MCP + Web UI + file watcher + wikiloop stdio ← subprocess MCP for embedded agents + +[Agent integration] + Claude Code / Cursor / any MCP client + → connects via http://localhost:8766/mcp (HTTP) + → or spawns wikiloop stdio (no network) + +[LLM for distillation] + Ollama (local) ← fully offline + or on-premise API endpoint +``` + +No data leaves the boundary. No external services required. The entire stack runs on a single machine or internal server. + +--- + +## Comparison with Cloud RAG in Regulated Industries + +| Factor | Cloud RAG | WikiLoop (local) | +|---|---|---| +| Data leaves premises | Yes (embedding API, LLM API) | No | +| Vendor lock-in | High (Pinecone, OpenAI, etc.) | None (plain files, SQLite) | +| Compliance audit | Hard (black-box vectors) | Easy (git diff, Markdown) | +| Air-gap capable | No | Yes | +| Infrastructure required | Vector DB, embedding service, LLM API | Single binary | +| Knowledge portability | Low | High (OKF-compatible directory) | +| Setup time | Days to weeks | Minutes | +| Cost at query time | Per-API-call | Zero | diff --git a/docs/site/zh-CN/ecosystem/use-cases.md b/docs/site/zh-CN/ecosystem/use-cases.md new file mode 100644 index 0000000..1c8150c --- /dev/null +++ b/docs/site/zh-CN/ecosystem/use-cases.md @@ -0,0 +1,213 @@ +# 行业使用场景 + +RAG 和 LLM Wiki 知识库在各行业的使用方式——存什么内容、谁在用、哪种方式更合适。 + +> **快速参考：** +> - **RAG** — 适合文档量大、更新频繁、需要精确引用来源的场景 +> - **LLM Wiki** — 适合稳定的结构化知识：概念、SOP、术语、决策记录 +> - **两者结合** — Wiki 提供稳定知识骨架，RAG 处理动态检索深度 + +--- + +## 律师事务所 + +**知识库存什么：** 案例法与判决先例、合同模板库、M&A 尽调历史文件、法规更新（SEC、GDPR 等）、内部 SOP 与卷宗管理手册。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 法律研究加速 | 助理律师、初级律师 | 自然语言提问，RAG 从案例库返回引用注脚 | +| 合同审查 | 资深律师 | 上传合同，系统对比历史条款、标记风险点 | +| M&A 尽调 | 并购团队 | 扫描数千份文件，提取关键负债与风险 | +| 合规监控 | 合规官 | 自动索引法规更新，推送相关变更提醒 | +| 客户自助门户 | 客户 | 常见法律问题由 AI 基于内部知识库自助回答 | + +**方案选择：** RAG 为主（文件量庞大、法规频繁更新、必须精确引用）。LLM Wiki 负责法律概念定义、内部术语表、标准流程说明。代表工具：Harvey AI、Thomson Reuters CoCounsel。 + +--- + +## 会计 / 财务事务所 + +**知识库存什么：** GAAP/IFRS/税法条文与解释性指南、IRS 出版物、审计准则（PCAOB/ISA）、客户财务报表与审计底稿、SEC/FASB 规则更新、标准参与信函模板。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 税务研究 | 税务经理、会计师 | 自然语言查询税法条文，获得引用注脚 | +| 审计支持 | 审计团队 | RAG 检索审计准则，生成审计备忘录草稿 | +| 财务报表分析 | 分析师、投顾 | 上传 10-K/10-Q，LLM 摘要关键风险和异常 | +| 监管变更跟踪 | 合规官 | 自动摄取 FASB/PCAOB 新规，推送影响提醒 | +| 客户理财顾问机器人 | 财富管理客户 | 基于客户投资组合回答专属问题 | + +**方案选择：** RAG 为主（税法/准则量大、版本更新快）。LLM Wiki 负责 GAAP 核心概念、内部操作手册、常见问答。代表工具：Microsoft Copilot for Finance、Workiva AI、Bloomberg Tax AI。 + +--- + +## 制造工厂 + +**知识库存什么：** 设备操作手册、维修服务公告、故障历史与维修日志、SOP、质量检验清单、OSHA/ISO 安全标准、供应商规格与 BOM。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 设备故障排查 | 现场技师 | 自然语言描述故障，RAG 返回手册对应步骤 | +| 质量管控 | QC 人员 | 查询历史缺陷模式，获取检验 SOP | +| 安全合规 | 安全主管 | 实时检索 OSHA/ISO 条款，应对现场问题 | +| 新员工入职培训 | 新工人 | 问答式学习工厂流程，减少对老员工依赖 | +| 供应链询盘 | 采购 | 查询供应商交期、零件规格与替代选项 | + +**方案选择：** RAG 为主（设备手册 PDF 量大、需按型号精确检索）。LLM Wiki 负责通用操作规范、术语词典、入职培训路径。两者结合：Wiki 作结构化知识骨架，RAG 检索具体手册细节。 + +--- + +## 软件开发团队 + +**知识库存什么：** 技术文档、API 参考、架构决策记录（ADR）、Confluence/Notion 页面、Jira 工单历史、Bug 报告与解决方案、代码注释与 README、关键技术讨论记录。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 开发者文档问答 | 工程师 | 自然语言查询 API 用法、架构惯例 | +| IT 帮助台自动化 | IT 支持 | 1/2 级工单由 AI 基于内部 KB 自动解答 | +| 新工程师 Onboarding | 新工程师 | 询问系统设计和代码结构，快速上手 | +| 代码审查辅助 | Reviewer | AI 检索类似历史 PR/决策，辅助评审 | +| 故障响应 | SRE/DevOps | 快速检索 runbook、历史事故处置记录 | + +**方案选择：** LLM Wiki 为主（架构设计、概念说明、规范文档——WikiLoop 的核心场景）。RAG 负责大量非结构化文档（Confluence、Jira、Slack 归档）。两者结合：WikiLoop 存稳定知识，RAG 动态检索最新 Issue/PR。代表工具：GitHub Copilot Enterprise、Glean、Guru。 + +--- + +## 客服中心 + +**知识库存什么：** 产品/服务 FAQ、退换货政策、服务协议、历史工单与解决案例、客服话术脚本与升级流程 SOP、CRM 客户历史记录。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 实时坐席辅助 | 客服代表 | 通话中实时弹出 RAG 建议答案，减少查找时间 | +| 智能机器人自助 | 终端客户 | 常见问题由 AI 直接解答，减少人工工单量 | +| 工单分类路由 | 系统自动 | AI 理解问题内容，智能分配给对应专家组 | +| 新客服培训 | 新员工 | 模拟问答，熟悉话术和产品知识 | +| 首次解决率优化 | 质检 | 分析未命中知识点，持续更新知识库 | + +**方案选择：** RAG 为主（客服知识更新频繁，促销政策、新产品实时变动）。LLM Wiki 负责稳定的标准话术、退货政策、升级流程。代表工具：Zendesk AI、Salesforce Einstein、Amazon Q。 + +--- + +## 医疗 / 医院 + +**知识库存什么：** EHR 摘要与患者历史、临床指南、药物相互作用数据库、ICD-10/CPT 编码规范、医院内部流程、护理 SOP、医学文献（PubMed 等）。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 临床决策支持 | 医生 | 查询药物禁忌、诊疗指南，辅助诊断 | +| 病历交接摘要 | 护士、住院医 | AI 自动摘要患者 EHR，加速晨交班 | +| 医疗编码辅助 | 编码员、财务 | RAG 从历史案例匹配正确 ICD/CPT 代码 | +| 患者 FAQ 机器人 | 患者 | 查询预约、费用、用药注意事项 | +| 影像/病理报告辅助 | 放射科、病理科 | AI 检索类似历史报告，辅助起草诊断 | +| 药物查询 | 药剂师、护士 | 实时检索药典与医院处方集 | + +**方案选择：** RAG 为主（医学文献量极大、临床指南更新快、必须有来源证据）。LLM Wiki 负责医院内部 SOP、护理流程、常见病种护理路径。**注意：** HIPAA 合规要求私有化部署。代表工具：Epic AI、Microsoft Azure Health Bot、Nuance DAX。 + +--- + +## 金融 / 投行 + +**知识库存什么：** 监管法规（Basel III、MiFID II、Dodd-Frank）、研报、财报、盈利电话会议记录、交易历史、风险模型文档、KYC/AML 合规流程、内部投资备忘录。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 投资研究合成 | 研究分析师 | RAG 整合多源研报/财报，生成摘要洞察 | +| 合规问答 | 合规官、风险官 | 查询最新监管条文，判断业务是否合规 | +| 客户顾问支持 | 理财顾问 | 通话中实时拉取产品信息与政策解读 | +| 反欺诈知识库 | 风控团队 | 跨团队共享新型欺诈模式识别知识 | +| M&A 尽调 | 投行 | 快速扫描目标公司财务与法律文件 | +| 交易员合规助手 | 交易员 | 交易前查询合规规则，避免违规操作 | + +**方案选择：** RAG 为主（监管文件量大、更新频繁）。LLM Wiki 负责内部产品知识、标准化合规流程。**注意：** 高度敏感数据需气隔离（Air-gapped）或私有云部署。 + +--- + +## 教育 / 培训机构 + +**知识库存什么：** 课程内容、教学材料、入学政策、奖学金 FAQ、考试大纲、历年真题解析、教师研究成果、认证合规文档。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 学生问答机器人 | 学生 | 自助查询课程安排、选课规则、奖学金 | +| 个性化学习助手 | 学生 | AI 基于课程 KB 提供答疑与辅导 | +| 教职员工知识库 | 教师、行政 | 查询院系政策、表格流程、合规要求 | +| 企业培训检索 | 培训员工 | 查询培训材料、考点、历史案例 | +| 毕业论文辅助 | 研究生 | 语义检索相关文献与已有研究结论 | + +**方案选择：** 两者结合——课程结构与教学路径写成 Wiki，具体题目和研究文献用 RAG 动态检索。LLM Wiki 为主适合 K12 标准化教材和企业固定培训内容。代表工具：Khanmigo、Microsoft Copilot for Education。 + +--- + +## 政府 / 公共机构 + +**知识库存什么：** 法律法规全文、政策文件、政府服务 FAQ（税务、社保、签证）、跨部门业务流程与审批手册、历史政策解读、内部规章制度。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 市民服务门户 | 普通市民 | 24×7 自助查询税务、社保、证件办理流程 | +| 政策文件检索 | 公务员 | 自然语言检索政策条文，辅助决策起草 | +| 跨部门知识共享 | 各部门 | 统一知识库打破信息孤岛 | +| 合规审查 | 审计官员 | 检索法规对照，辅助合规审查工作 | +| 议会问答支持 | 政务人员 | 快速检索历史文件为质询准备答辩材料 | + +**方案选择：** 两者结合——稳定政策法规写成结构化 Wiki 供市民理解，具体条文全文索引用 RAG。**注意：** 政府数据主权要求本地化部署。代表工具：Palantir AI Platform、Microsoft Azure Government。 + +--- + +## 电商 / 零售 + +**知识库存什么：** 商品详情、规格参数、使用说明、退换货政策、物流追踪 FAQ、历史工单、库存/供应链/供应商数据、促销活动规则与会员权益。 + +| 使用场景 | 使用者 | 具体做法 | +|---|---|---| +| 智能商品问答 | 客户、客服 | 自然语言查询商品参数、库存、发货时效 | +| 退换货自动化 | 客户 | AI 自助完成退货申请，减少人工处理 | +| 个性化推荐 | 购物者 | 知识库结合用户画像，精准推荐商品 | +| 供应链 Q&A | 采购、运营 | 查询库存、供应商交期、备货建议 | +| 坐席辅助 | 客服 | 通话中实时弹出政策答案，减少处理时长 | + +**方案选择：** RAG 为主（SKU 量极大、商品信息频繁变动）。LLM Wiki 负责通用退货政策、会员等级规则、FAQ。代表工具：Salesforce Einstein、Zendesk AI、Shopify Sidekick。 + +--- + +## 更多行业 + +| 行业 | 知识库内容 | 核心使用场景 | 推荐方案 | +|---|---|---|---| +| **电信** | 网络故障手册、套餐对比、流失挽留预案 | 技术人员故障排查、客服推荐套餐、挽留话术 | RAG 处理故障，Wiki 存话术 | +| **保险** | 保单条款、理赔 SOP、核保标准、监管法规 | 理赔员查询条款、核保辅助决策、合规问答 | RAG 为主，Wiki 存稳定流程 | +| **生命科学 / 制药** | 临床试验数据、FDA/EMA 法规、药物研发文档 | 研发人员文献检索、监管申报文件起草 | RAG + Wiki 结合 | +| **房地产** | 房源信息、市场报告、合同模板、地方法规 | 经纪人 Q&A、市场分析、合同审查 | RAG 为主 | +| **媒体 / 出版** | 存档文章、编辑规范、版权数据库 | 记者调研、编辑一致性检查、版权核查 | RAG + Wiki 结合 | + +--- + +## WikiLoop 在各行业的切入点 + +WikiLoop 本地优先、MCP 原生、纯 FTS 的特性，在以下行业场景特别契合： + +| 行业 | 契合度 | 具体用法 | +|---|---|---| +| **律师事务所** | 高 | 将案例研究笔记、决策记录、客户案件摘要编译成可审计的 Wiki | +| **会计事务所** | 高 | GAAP 概念结构化知识库、内部业务方法论 Wiki | +| **制造工厂** | 中 | 工厂 SOP、设备词典、入职知识库——仅适合稳定内容 | +| **软件开发** | 非常高 | ADR 库、架构知识库、团队 Onboarding Wiki——WikiLoop 的核心场景 | +| **客服中心** | 中 | 标准话术 Wiki、稳定政策参考——与 RAG 结合处理动态产品数据 | +| **医疗机构** | 高 | 内部临床 SOP、护理路径 Wiki——私有化部署满足 HIPAA 要求 | +| **金融机构** | 高 | 内部合规流程 Wiki、产品知识库——气隔离部署满足安全要求 | +| **教育机构** | 高 | 课程结构 Wiki、机构知识库、研究综合 | +| **政府机构** | 高 | 政策解读 Wiki、市民 FAQ 知识库——本地部署满足数据主权要求 | +| **研究团队** | 非常高 | 文献综合、概念图谱、决策记录——WikiLoop 最核心的使用场景 | + +--- + +## 选型决策框架 + +| 维度 | 选 RAG | 选 LLM Wiki | 两者结合 | +|---|---|---|---| +| 文档量 | 数万页以上 | 几百页以内 | 大小混合 | +| 更新频率 | 每天/每周更新 | 每月/每季更新 | 混合节奏 | +| 精确引用需求 | 必须（法律/医疗） | 一般性即可 | 敏感场景 | +| 内容结构化程度 | 低（PDF、邮件） | 高（流程、概念） | 混合格式 | +| 用户问题类型 | 开放式检索 | 固定模式问答 | 两类都有 | diff --git a/docs/site/zh-CN/ecosystem/wikiloop-industries.md b/docs/site/zh-CN/ecosystem/wikiloop-industries.md new file mode 100644 index 0000000..eb3b725 --- /dev/null +++ b/docs/site/zh-CN/ecosystem/wikiloop-industries.md @@ -0,0 +1,186 @@ +# WikiLoop 行业应用 + +WikiLoop 是一款面向 AI Agent 的本地优先知识搜索引擎。其架构——单一二进制文件、SQLite FTS5、纯 Markdown、MCP 协议——使其在隐私要求严格、合规要求高、离网环境下具有独特优势。 + +## 为什么 WikiLoop 适合高隐私行业 + +大多数 RAG 系统需要外部基础设施：云端 Embedding API、向量数据库、托管 SaaS 服务。WikiLoop 消除了所有这些依赖。 + +| 需求 | WikiLoop 的满足方式 | +|---|---| +| **数据不出本地** | 所有数据存于本地，搜索和索引零外部 API 调用 | +| **无需 Embedding 模型** | 纯 SQLite FTS5 + BM25，无向量模型、无 Embedding API、无 GPU | +| **零外部基础设施** | 单一二进制文件 + 一个 SQLite 文件，无需 Redis、Kafka、Postgres 或向量数据库 | +| **气隔离（Air-gap）兼容** | stdio MCP 模式作为本地子进程运行，零网络访问即可工作 | +| **知识可审计** | 所有 Wiki 页面是 git 中的纯 Markdown，每次变更都是 diff | +| **无供应商锁定** | OKF v0.1 兼容，KB 就是一个文件目录，可迁移到任何系统 | +| **支持本地 LLM** | 提炼功能支持 Ollama 或任意本地模型，文档从不发送到 OpenAI | +| **BYOK（自带密钥）** | 如使用云端 LLM 提炼，只需提供 API Key，无需 SaaS 订阅 | + +--- + +## 律师事务所 + +**隐私关切：** 律师-客户特权。客户文件不能离开律所基础设施。 + +**WikiLoop 的使用方式：** + +- 将案例研究笔记、法律备忘录、先例分析编译为结构化 Wiki 页面 +- 构建全所 ADR 风格的决策库：历史交易结构、诉讼策略、和解模式 +- 每个案件对应一个 source-note，法律理论、监管框架合成为 concept 页面 +- `wikiloop lint` 验证引用完整性——每个主张都可追溯到原始文件 +- Git 历史记录提供完整审计链，支持特权审查 + +**部署方式：** 律所自管服务器或笔记本电脑上的单一二进制文件，初始设置后无需联网。LLM 提炼可使用本地模型（Ollama + Llama 3）或内部 API。 + +--- + +## 会计 / 财务事务所 + +**隐私关切：** 客户财务数据受 NDA 保护，受 SOX、GDPR 等数据处理限制。 + +**WikiLoop 的使用方式：** + +- 结构化 GAAP/IFRS 概念库：每个准则对应一个 Wiki 页面，通过 `related_to` 链接关联相关解释 +- 内部业务方法论 Wiki——事务所处理特定审计类型的方法，以 decision 页面形式记录 +- 将客户专属知识（脱敏后）编译为跨项目学习库，不暴露原始客户数据 +- 知识空白分析（`wikiloop synthesize --gaps`）在旺季前发现文档不足的领域 + +**部署方式：** 内部部署或私有云。使用本地模型提炼，客户数据完全留在事务所边界内。 + +--- + +## 医疗机构 / 医院 + +**隐私关切：** HIPAA（美国）、GDPR（欧盟）及各地健康数据法规，患者数据不能触及公共云。 + +**WikiLoop 的使用方式：** + +- 内部临床 SOP Wiki：护理规程、护理路径、术后流程——结构化且可搜索 +- 药物处方集知识库：医院专属处方集从标准参考文献编译而成，由药剂科团队持续维护 +- 事故与未遂事件知识库：脱敏案例摘要提炼为 source-note，根本原因模式合成为 concept 页面 +- 新住院医师入职：机构专属流程和规程存入可搜索 Wiki，减少对高年资医师的依赖 +- 研究团队知识库：文献综合、试验设计决策、IRB 批准的方案文档 + +**部署方式：** 气隔离医院内网，`wikiloop stdio` 作为临床 AI 工具的子进程运行，零网络出口。 + +--- + +## 金融机构 / 投行 + +**隐私关切：** 重大非公开信息（MNPI）、交易保密、监管合规（MiFID II、SEC Rule 10b-5）。 + +**WikiLoop 的使用方式：** + +- 内部合规流程 Wiki：已批准交易实践、升级流程、预先审批工作流的结构化文档 +- 交易知识库：每笔完成交易对应一个 source-note，记录交易结构、关键决策和理由——在跨团队之间积累机构记忆 +- 法规解释库：合规团队维护全所对特定监管规则的解释 Wiki，`contradicts` 链接标记冲突解释 +- 风险模型文档：量化分析师将模型假设、局限性和验证结论记录为 decision 页面 + +**部署方式：** 气隔离交易大厅基础设施或私有云，无公网路径。提炼使用内部 LLM。 + +--- + +## 制造业 / 工业 + +**隐私关切：** 商业机密、专有生产工艺、设备设计中的知识产权。 + +**WikiLoop 的使用方式：** + +- 工厂 SOP Wiki：每条产线的标准操作规程，可按工序步骤、设备类型或产品系列搜索 +- 设备故障排查知识库：从维护日志、服务公告和技师现场记录提炼——Agent 维修时实时查询 +- 质量缺陷模式库：历史缺陷数据编译为 source-note，反复出现的根本原因合成为 concept 页面 +- 新员工入职培训：结构化知识库减少工艺知识转移对老员工的依赖 +- 供应商与物料知识：规格书、替代品历史、供应商可靠性记录编译为 Wiki 页面 + +**部署方式：** 工厂 OT 网络，通常与互联网物理隔离。WikiLoop 运行在本地服务器；技师通过 AI 助手或自定义对话界面经 MCP 查询。 + +--- + +## 政府与国防 + +**隐私关切：** 国家安全、涉密信息、数据主权要求。 + +**WikiLoop 的使用方式：** + +- 政策解读 Wiki：特定法规在实践中如何适用的结构化文档，按政策领域可搜索 +- 市民服务知识库：FAQ 和办事指南，由主题专家维护 +- 跨部门知识共享：通用操作规程编译为共享 Wiki，减少跨部门重复建设 +- 采购与合规知识：历史采购决策以 decision 页面形式记录，附理由和备选方案 + +**部署方式：** 主权云或内部部署，无外部连接。`wikiloop serve` 部署在内部网络边界内。stdio MCP 供嵌入式 Agent 使用。OKF 兼容格式支持在安全系统间传输知识。 + +--- + +## 能源与公用事业（核电、油气） + +**隐私关切：** 关键基础设施保护、专有运营数据、物理气隔离要求。 + +**WikiLoop 的使用方式：** + +- 运维知识库：设备手册、事故报告、流程更新提炼为结构化 Wiki 页面 +- 安全与监管合规 Wiki：电站对 NRC/HSE 法规的专属解释、应急流程、安全案例文档 +- 老龄化员工知识留存：经验丰富的工程师在退休前将隐性知识捕获为 source-note 和 concept 页面 +- 事故学习库：事后报告提炼并交叉关联——事故响应时 `related_to` 边自动浮现类似历史事件 + +**部署方式：** 物理气隔离的运营技术（OT）网络。WikiLoop 零外部依赖运行。提炼离线批量处理；索引实时响应 Agent 查询。 + +--- + +## 研究机构与智库 + +**隐私关切：** 发表前研究保密、经费合规、机构知识产权。 + +**WikiLoop 的使用方式：** + +- 文献综合：研究人员将论文放入 `raw/`，WikiLoop 提炼 source-note，并跨语料合成 concept 和 comparison 页面 +- 研究决策记录：方法论选择、假设修订、数据来源评估以 decision 页面形式记录 +- 跨项目知识共享：通用方法论和发现编译为共享机构 Wiki +- 知识空白分析：`wikiloop synthesize --gaps` 在撰写经费申请前识别研究不足的领域 + +**WikiLoop 在此场景的特别优势：** 这最接近 WikiLoop 的原始设计意图。Agent 迭代搜索知识库、跟随论文间的 `related` 链接、自行综合结论——正是研究人员的工作方式。 + +**部署方式：** 研究人员自己的电脑或机构服务器，无需云端。本地 Ollama 模型提炼，确保发表前数据完全离线。 + +--- + +## 私有化部署架构 + +以上所有行业，典型的 WikiLoop 部署如下： + +``` +[本地 KB 目录] + raw/ ← 放入原始文档 + wiki/ ← 提炼后的页面（Markdown + git） + index/ ← SQLite FTS 索引（自动管理） + +[WikiLoop 二进制] + wikiloop serve ← HTTP MCP + Web UI + 文件监听 + wikiloop stdio ← 嵌入式 Agent 子进程 MCP（无网络） + +[Agent 接入] + Claude Code / Cursor / 任意 MCP 客户端 + → HTTP 连接 http://localhost:8766/mcp + → 或 stdio 启动 wikiloop（无网络） + +[提炼用 LLM] + Ollama（本地）← 完全离线 + 或内部 API 端点 +``` + +数据不出边界。无需外部服务。整个技术栈运行在单台机器或内部服务器上。 + +--- + +## 与云端 RAG 在合规行业的对比 + +| 因素 | 云端 RAG | WikiLoop（本地） | +|---|---|---| +| 数据是否出境 | 是（Embedding API、LLM API） | 否 | +| 供应商锁定 | 高（Pinecone、OpenAI 等） | 无（纯文件、SQLite） | +| 合规审计难度 | 高（黑盒向量） | 低（git diff、Markdown） | +| 气隔离支持 | 否 | 是 | +| 所需基础设施 | 向量数据库、Embedding 服务、LLM API | 单一二进制文件 | +| 知识可移植性 | 低 | 高（OKF 兼容目录） | +| 部署时间 | 数天到数周 | 几分钟 | +| 每次查询成本 | 按 API 调用计费 | 零 |