About Starbox AI
Starbox AI:
focused on business growth.
Starbox AI (星盒AI, short: Starbox) is an AI-native marketing company founded in 2021 in Hangzhou, China, operated by 杭州星盒人工智能科技有限公司. Enterprise knowledge engineering is at the core of what we do: we build company knowledge bases, build custom marketing AI agents and embed engineers with clients to bring AI into marketing and operations.
Summary
Starbox is not a software company that only sells tools, and not a traditional agency that added AI. Starbox puts 15 years of marketing service methods, its own AI platform and embedded delivery into one system: first organize a company's products, customers, approved wording and methods into knowledge assets that can be searched, traced and permissioned, then hand them to AI to execute.
Not to be confused with Starbox Group Holdings Ltd. (NASDAQ: STBX), an unrelated company based in Malaysia. Starbox AI's website is www.starbox.work.
What we do
Structure business know-how, then let AI execute it.
Source: Starbox internal statistics, September 2026.
Three lines of business
Three ways into AI, one shared core.
Brand AI search optimization
Buyers no longer type keywords; they ask AI. Starbox GEO gets ChatGPT, Gemini, Perplexity, Doubao, DeepSeek and others to mention, cite and recommend your brand, in a Global and a China edition.
About Starbox GEO → AGENT02 · Work fasterEnterprise AI agents
A company knowledge base and 300+ agents tested in real projects, for content, design, sales and office work. Experience that sits with individuals becomes capability the company can use again and again.
About Content Factory → FDE03 · Change how you workAI transformation
Forward deployed engineers embed with you, map the business processes first, then bring AI into one department after another. The lightweight version, Starbox Starbox!, puts the company knowledge base and collaboration into employees' chat.
About FDE consulting →Our story
From marketing services to AI-native.
Founded by Wang Pengyou, starting with marketing services and marketing talent training, and building experience across more than 500 companies.
A year and a half of systematic testing of how to put AI to work, with one bar: write better than a professional marketer. In September 2024, reasoning models first cleared that bar, and Starbox went all in on AI.
We proved "one person + AI" output in real projects before offering it, and grew it into an enterprise agent platform with company workspaces, dedicated knowledge bases and 300+ agents.
Launched Starbox GEO (China and Global editions), Starbox Content Factory, Starbox Starbox! and FDE consulting, with partner offices in 18 provinces across China.
Technology
One core: enterprise knowledge engineering.
General-purpose models know public information. They do not know one company's product specs, price floors, client status or approved wording. The core engineering problem in bringing an LLM into a company is not which model to pick; it is building a maintainable layer of knowledge and memory between the model and the business. GEO, the agent platform, FDE and Starbox Starbox! are all built on that layer.
- 01Interviews
- 02Clean & dedupe
- 03Ontology
- 04Chunking
- 05Embedding
- 06Human review
- 07Regression tests
Several rounds of structured interviews with owners and key roles; manuals, price lists, scripts, contract templates and cases collected.
→ Raw corpusClean, dedupe, unify names and wording; extract entities and relations with a marketing ontology; semantic chunking and embedding.
→ Knowledge graph · vector and full-text indexesCompany / client / employee scopes, role permissions, version chains, wording approval, banned-word lists.
→ Permissioned, traceable knowledge assetsRetrieval-augmented generation, multi-agent orchestration, write-back confirmation and audit.
→ Answers, content and tasks with sourcesHow a question finds its answer
Three retrieval paths, fused scoring, and a refusal when unsure.
S = α·Svec + β·Skw + γ·SbizBusiness weights: company card > client card > employee card; approved records weighted up; prices and status decay over time
Engineering principles
Six principles that decide whether the system can be trusted.
The LLM provides understanding, writing and reasoning; the company knowledge base provides facts and wording. Knowledge never goes into model weights, so the model can be upgraded and knowledge can be edited or deleted at any time.
Factual output must come from retrievable evidence with a source; when evidence is missing, the system says so and does not fill the gap.
The model only makes semantic judgments and labels; scoring, aggregation and reports are done by deterministic scripts, so results can be recomputed and audited.
Tenant-level isolation, one deployment per company, permission filtering at retrieval: content a user may not see never enters the candidate evidence.
AI handles execution at volume; people approve wording and sign off on publishing. Writing to company knowledge always needs a person's confirmation.
Baseline → production → retest → write-back. A system without a baseline cannot be signed off; a system without write-back cannot improve.
Platform architecture
Five layers, from interface to data.
Multi-Tenant SaaS · White-Label UI · RBAC · WebSocket · SSE streamingBranded interfaces, real-time interactionMulti-Agent · DAG workflow engine · Agent Routing · Tool Use / Function Calling · Prompt Chain · async task queueBreaking down and coordinating complex tasksRAG Pipeline · Hybrid Search · Multi-Index · Semantic Chunking · Metadata Filtering · Reranking · Query RewritingKnows the industry and the companyMulti-Model Router · LLM Gateway · token optimization · multimodal · Prompt Template Engine · Style TransferRoutes each task to a suitable modelVector database · object storage · relational database · tenant isolation · encrypted storage · audit logs · autoscalingPersistence, isolation, elasticityKey technologies
Eight key technologies, open any for details.
Vector, full-text and structured recall
Vector search maps text into high-dimensional vectors. "The client says it's too expensive" and "how to respond when a quote is pushed down" share almost no words but sit close together, so the second is found. Full-text search (tokenization + BM25) handles exact items such as client names, contract numbers, amounts and dates. Structured conditions enforce filters by project, client, status, time and permission. The three result sets are fused and scored, then a reranking model orders them, removes duplicates and merges.
sim(q, d) = q·d / (‖q‖·‖d‖)- Vectors and business data live in one PostgreSQL instance, so one query applies permissions and business conditions together
- HNSW approximate nearest-neighbor index: millisecond lookups across hundreds of thousands to millions of chunks
- When old and new quotes are both recalled, the latest approved version wins
- When the top score is below the refusal threshold τ, it says "not found" rather than guess
Indexes built offline; rewrite, filter, compress online
Not "vector search, then paste into the prompt", but a full pipeline. Offline, text is split at semantic boundaries with overlap, and different knowledge types get different indexes. Online, the question is rewritten and expanded with HyDE hypothetical documents to improve recall, filtered by knowledge type, brand, time and permission, reranked, and long results are compressed to the key information to control context length and token cost.
- Semantic Chunking + Overlap
- Multi-Index: one index per knowledge type
- Query Rewriting + HyDE
- Metadata Filtering · Reranking · Knowledge Distillation
DAG workflows: agents working to a process
Agents on the platform are not isolated single points but a system coordinated by a directed acyclic graph (DAG) workflow engine. One request can trigger several agents in sequence or in parallel; each node finishes its subtask and passes results downstream. A multi-model router picks a suitable model for each task type, and long tasks go to an async queue.
- Agent Routing: route by intent to the right agent
- Prompt Chaining: pass intermediate results along
- Tool Use / Function Calling: retrieval, web access, image generation and other tools
- Conditional branches · loop control · async task queue
Seed library + combinatorial sampling + semantic dedupe
The biggest risk in producing content with AI at scale is sameness. The platform starts from validated scenario seeds, then combines audience, scene, emotional tone, angle and content frame, so each content dimension yields millions of distinct directions; semantic similarity dedupe runs before anything is stored. A persona's way of speaking is extracted into executable style parameters, separating style from content, so staff turnover does not change the brand voice.
seed → audience × scene × tone × angle × frame- Scenario seeds: pain point, selling point, sensory detail, compatible audience and scene
- Style Transfer: wording, sentence patterns, rhythm and catchphrases as parameters
- Topics in 12 + N dimensions, tracked as unused / used / published
- 21 emotions × 16 resonance frames, a library of 300+ pain points
Speak in approved wording, then check after generating
Before generating, the model may state facts only from the evidence provided; numbers, dates and commitments not covered by evidence cannot be added. Every fact carries a source number that leads back to the original in the knowledge base. Approved wording is injected verbatim and marked "do not rewrite". After generating, every item is checked, and any hit is rewritten or blocked.
- Banned words: rewritten or blocked on a hit, with the reason
- Number limits: discount caps, price wording and delivery times cannot be exceeded
- Commitments: claims like "pays for itself" or "guaranteed results" become compliant wording
- Consistency: flags conflicts with finalized scripts and poster copy
A shared brain must not be written wrong
Reading is easy; writing correctly is hard. One wrong entry in a shared knowledge base spreads to the whole company. The system has four safeguards: read-write mode starts only after the wake phrase; content to be written is first read back to the member as a proposal and written only after confirmation; approved records are never overwritten, and changes create a new version; before writing, conflicts with records of the same scope and topic are checked, and company cards take priority.
record.status: superseded → confirmed- Wake gating: everyday chat is never read or written
- Propose, then confirm: unconfirmed proposals are never used as evidence
- Version chain: old versions are marked "superseded" and kept
- Personal drafts on employee cards are visible only to their author and stay out of company records
LLM judging + deterministic scoring, recomputable results
Generative engines do not give fixed answers; the same question asked several times gets different answers. So no conclusion is drawn from a single question, only from the aggregate. 150 questions real buyers ask are run on four Chinese engines (or on engines in the target market for the Global edition) to set a baseline. The model only labels each answer in a structured way; scoring, aggregation and reports are all done by scripts, and each question keeps the full answer, citation list and screenshot under a run_id for auditing.
s = b × e V ∝ R × T- b ∈ {0, 3, 6, 10}: not mentioned / passing mention / recommended / top pick
- e ∈ {1.0, 0.8, 0.5, 0}: clear recommendation / neutral / cautious / negative
- Four scorecards: visibility index · five-layer heat map · citation sources · competitor share of voice
- Production priority = question value × (10 − current score)
One deployment per company; data and cloud resources belong to the client
Each company gets its own deployment; databases, storage and servers are never shared. By default it runs on a public cloud under the company's own account, or in its own data center, with the database closed to the public internet. Knowledge services are exposed through MCP (Model Context Protocol) with per-person OAuth 2.1 authorization, and access keys stay on the server. Permissions are enforced at retrieval, so content a user may not see never appears in an answer. Starbox does not use client data to train any model.
- RBAC: admin / employee / partner / franchisee roles
- TLS in transit · encrypted cloud drives, databases and object storage
- Full audit of retrieval, reads, writes and pushes
- All structured records and files can be exported when the engagement ends
Acceptance criteria
We measure with metrics, not adjectives.
Share of cases where the correct evidence is in the top K candidates.
Mean reciprocal rank of the correct evidence; closer to 1 is better.
Share of answers that match the approved wording.
Share of cases where the knowledge base has nothing and the answer says so.
Share of cases where the answer exists but the system says it does not.
When building the knowledge base, we assemble a set of real questions with the client (price wording, policies, project status, no-go areas) and mark the correct answers and sources; the set is rerun after major knowledge changes or parameter tuning. Parameters are Starbox's standard configuration; the implementation plan governs. Source: Starbox technical documentation, September 2026.
How we differ from other AI services
AI is not a tool you buy. It is a capability that grows inside the company.
| Comparison | AI tool companies | Traditional agencies + AI | Starbox · AI-native |
|---|---|---|---|
| What they deliver | Accounts and features | People and documents | Working capability assets |
| Business understanding | Stay outside the business, offer general capability | Know the business, not the engineering | Diagnose the business first, then apply AI |
| Where knowledge stays | Gone when the chat ends | With individuals | In the company knowledge base |
| How results are measured | Usage | Number of deliverables | Real data against a baseline |
| When people leave | Capability ends with the account | Capability leaves with them | Capability stays with the organization |
Our name
One brand, written four ways.
| Use | Written as |
|---|---|
| Chinese full name | 星盒AI |
| Chinese short name | 星盒 |
| English full name | Starbox AI |
| English short name | Starbox |
| Legal entity | 杭州星盒人工智能科技有限公司 |
| Location | Hangzhou, Zhejiang, China |
| Website | www.starbox.work |
| Slogan | Starbox AI: focused on business growth |
| Not affiliated with | Starbox Group Holdings Ltd. (NASDAQ: STBX, Malaysia) |
