The short answer: an enterprise AI knowledge base is not a token bill. The real budget is driven by document parsing and cleanup, chunking and embeddings, access control, system integration, evaluation and ongoing operations. At 300,000 questions a year, model and embedding fees can be only a few thousand RMB — while a usable departmental project often costs several hundred thousand RMB. The gap is data and engineering, not the model list price.
This guide is for teams planning an internal knowledge base, a support knowledge base, or a technical-document assistant. Data cut-off: 25 October 2026. Prices come from public DeepSeek, Alibaba Cloud Model Studio and Zilliz pages; budget ranges are editorial estimates, not quotes, and we do not promise answer accuracy or a delivery date. Infrastructure is covered separately in What private AI deployment really costs.
Cheap tokens do not make a cheap knowledge base. Messier documents and access rules cost more. Without an evaluation set, a demo is not production.
What does an enterprise AI knowledge base actually solve?
It suits document-heavy, frequently changing, permission-sensitive questions where answers need citations. It does not turn contradictory or outdated documents into truth, and it cannot replace a document owner.
RAG retrieves relevant passages, sends them to a model, and asks the model to answer from those passages. That works when knowledge changes, sources matter and different roles see different documents. It fails when the corpus has no owner, no version control and no access labels.
| Scenario | Good fit | Not a fit / fix first |
|---|---|---|
| Internal policies and processes | Many documents, frequent updates, repeated questions | Conflicting versions, no effective date or owner |
| Support and after-sales | Product docs, known errors and tickets can be structured | No ticket closure, no traceable answers, no role-based access |
| R&D and process documents | Standards, drawing notes and past issues are searchable | Drawings or formulas cannot be parsed; sensitive data cannot leave the intranet |
| Cross-system Q&A | Knowledge base plus read-only OA/CRM/ERP interfaces | Direct writes into core systems with unclear responsibility |
Editorial guidance. A knowledge base is not “upload every file”; document ownership, versioning, permissions and evaluation are part of the product.
Where does the money go?
Most of the budget goes into parsing, chunking and labelling, permissions, integration, evaluation and operations. Model APIs are only one line item; comparing token prices alone underestimates the project.
| Cost bucket | One-off / annual | What it includes | What teams miss |
|---|---|---|---|
| Document access and parsing | Mostly one-off | PDF/Word/Excel/scans/web pages; tables, images, OCR, deduplication | Bad parsing cannot be fixed by a better retriever |
| Cleanup, chunking, labels | One-off + ongoing | Chunk size, overlap, metadata, glossary, ACL tags, version cleanup | No metadata or ACL means retrieval crosses data boundaries |
| Embeddings and index | One-off + re-index | Text vectors, batch jobs, dimensions, index build | Changing the embedding model means re-embedding the corpus |
| Vector store and retrieval | Annual | Managed vector DB or self-hosted, replicas, backup, monitoring, SLA | Free tiers are fine for tests, not necessarily for production SLAs |
| Generation and reranking | Usage-based | Input/output tokens, long context, rerank, caching | Counting only one answer’s tokens, ignoring rerank and long context |
| Permissions, audit, integration | One-off + annual | SSO, document ACLs, audit logs, OA/IM/CRM/ERP interfaces, citations | Permission leakage, departed accounts, cross-team data exposure |
| Evaluation and operations | One-off + annual | Question sets, human review, bad cases, updates, training | No evaluation set means “it sounds right” is the only test |
Editorial cost map, not a vendor quote. The same “knowledge base” label can hide very different document volumes, permissions and integrations.
What do 300,000 questions a year cost in model and vector fees?
With 100,000 chunks, 300,000 questions a year and current DeepSeek Flash plus Alibaba text-embedding pricing, first-year model and embedding fees are roughly RMB 1,225–2,425 — excluding rerank, OCR, storage and people.
Assumptions: 100,000 chunks × 500 tokens = 50 million embedding tokens. 1,000 questions/day × 300 days = 300,000 questions; 2,000 input tokens and 500 output tokens per question. Embedding is a one-off build cost in this example.
| Item | Volume | Public price (25 Oct 2026) | Cost |
|---|---|---|---|
| Embedding | 100k chunks × 500 = 50M tokens | Alibaba text-embedding-v4: RMB 0.5 per 1M input tokens (Beijing) | ≈ RMB 25 one-off |
| Input tokens | 300k × 2,000 = 600M tokens | DeepSeek Flash: RMB 1–2 per 1M input tokens (off-peak/peak) | RMB 600–1,200 |
| Output tokens | 300k × 500 = 150M tokens | DeepSeek Flash: RMB 4–8 per 1M output tokens | RMB 600–1,200 |
| Total | — | — | ≈ RMB 1,225–2,425 first year |
Excludes rerank, OCR, storage, network, cache hits, retries, tax and discounts. Prices change; use the official pages before budgeting.
The counter-intuitive part is that generating 300,000 answers can cost less than one month of a content owner’s time. The hard part is getting the right passage, with the right permission, in front of the model. Teams that budget only tokens usually overrun during document cleanup and integration.
Why is the total budget so much higher?
Because parsing, permissions, integration, evaluation and operations dominate. A departmental single knowledge base commonly costs RMB 220,000–650,000 in the first year; token fees are a rounding item.
Example: 5,000 documents, 100,000 chunks, 300,000 questions/year, OA/IM integration and document-level permissions. The ranges below are editorial estimates, not quotes.
| Budget line | Low | High | Assumption |
|---|---|---|---|
| Parsing / cleanup / chunking | RMB 30,000 | RMB 100,000 | Clean documents → scans, tables and duplicate versions |
| Permissions / integration / UI | RMB 80,000 | RMB 250,000 | SSO, OA/IM, citations, audit |
| Evaluation / labelling | RMB 20,000 | RMB 50,000 | Real question set, bad-case review, acceptance |
| Vector store / infrastructure | RMB 10,000/yr | RMB 50,000/yr | Managed or self-hosted, backup and monitoring |
| Model / embeddings | RMB 1,000/yr | RMB 3,000/yr | 300,000-question example above |
| Operations | RMB 80,000/yr | RMB 200,000/yr | Sync, re-index, bad cases, permissions, training |
| First-year total | ≈ RMB 220,000 | ≈ RMB 650,000 | Departmental single knowledge base; excludes private hardware and full ERP work |
Editorial estimate, not a quote. Low assumes clean docs, existing cloud and light integration; high assumes scans, complex ACLs and multiple systems. Private or group-wide programmes cost more.
RAG, long context or fine-tuning?
Use RAG for changing knowledge, citations and permissions; long context for small, ad-hoc analysis; fine-tuning for stable style or behaviour, not for remembering facts.
| Approach | Best for | Poor fit | Cost profile |
|---|---|---|---|
| RAG | Changing knowledge, citations, layered access, many documents | Tiny corpus, simple questions, no document governance | Data, integration and operations dominate; tokens are small |
| Long context | Small document sets, one-off analysis, fast validation | Large corpus, frequent Q&A, cost-sensitive workloads | Every call pays for context; longer prompts cost more |
| Fine-tuning | Stable tone, format or domain behaviour | Fresh facts that change frequently | Labels, training and evaluation; updates are slower |
They can be combined: RAG for facts, fine-tuning for style. Do not use fine-tuning to replace permissions or knowledge updates.
How should you accept a RAG system?
Check retrieval recall, answer faithfulness, citation accuracy, out-of-scope refusal, permission isolation, latency and cost per answer. Answer accuracy alone is not enough.
| Metric | Definition | What to verify |
|---|---|---|
| Retrieval recall | Whether the right passage appears in Top-K | Break down by question type; fix parsing/chunking first |
| Answer faithfulness | Answer is supported only by retrieved passages | Require citations; say “unknown” when evidence is missing |
| Citation accuracy | Cited passage actually supports the claim | Sample and compare answer with source text |
| Out-of-scope refusal | Questions outside the corpus are declined | Prepare out-of-scope, sensitive and permission-denied questions |
| Permission isolation | Users only retrieve documents they may see | Test with different roles; keep audit logs |
| Latency and availability | P95 response, failure rate, retrieval timeout | Test real concurrency, not a single demo |
| Stale-answer rate | Answers citing old or invalid documents | Effective dates, versions and incremental sync |
| Cost per answer | Model + retrieval + human review + operations ÷ successful answers | Use a business denominator, not only tokens |
Thresholds must be agreed per scenario; there is no universal “95% is good”. Huawei Cloud’s RAG evaluation guidance also highlights hallucination, weak refusal and retrieval noise; prompts do not replace evaluation and permissions.
How should you run the POC and structure the contract?
Run a read-only shadow POC on real documents and real questions, validate parsing, retrieval, citations and permissions, then decide; write data ownership, ACLs, evaluation, model/vector replacement, SLA and exit into the contract.
- Scope: one department, one document class, one real question set. Do not ingest the whole company at once.
- Parse first: sample PDF tables, scans, versions and ACL labels. Parsing errors become retrieval errors.
- Read-only shadow: answer and cite first; do not write back into OA/ERP. Record sources and human review.
- Evaluate before launch: freeze a question set covering frequent, rare, out-of-scope and permission-denied cases.
- Contract: data/model/vector ownership, document ACLs, incremental sync, model and vector portability, SLA, audit and exit export.
- Pay by stage: parsing acceptance, POC, launch and stable operation. “Demo accepted” is not acceptance.
Is LingHarness a fit?
It fits enterprises that need private model access, permission boundaries and agent execution. Teams that only need a small document Q&A should start with a lightweight RAG product or POC.
LingHarness is positioned as an enterprise AI agent runtime and model-access foundation: private deployment, model routing, tool calls and observable execution. It is not a document parser, embedding model or vector database. Those components need to be selected separately; this article does not count third-party RAG components as Lingyi’s own technology.
If you only need a small document Q&A, a SaaS or open-source stack is usually faster and cheaper. If you have multiple systems, complex permissions, a private-data boundary and long-term operations, a unified AI foundation may be justified. For the platform comparison, see private AI deployment costs; for the Chinese version of this article, see 企业 AI 知识库要多少钱.
Need a knowledge-base cost model, POC scope or permission/evaluation design? Talk to us. We do not promise a fixed accuracy rate, delivery date or answer quality.
FAQ
Does an enterprise knowledge base always need a vector database?
No, but most semantic retrieval scenarios use a vector index. Small corpora can start with keyword/full-text or hybrid retrieval; introduce a vector store when semantic recall and scale justify the operational cost. “Vector database” is not a synonym for “intelligent”.
Can we build RAG on one million documents?
Technically yes, but cost and difficulty rise with volume, formats, permissions and update frequency. Fix parsing, deduplication, ACLs, incremental sync and evaluation first. More documents without those controls simply amplify errors; use layered knowledge bases and phased rollout.
Why do token fees look so low when the project is expensive?
Tokens only cover model and embedding calls. Most of the budget is document parsing/cleanup, permissions, integration, evaluation, UI and operations. Messier documents and more systems mean more human work; low token prices do not imply a low total cost.
Is private deployment always more expensive than SaaS?
Not always, but it adds servers, model deployment, backup, security and operations. Private deployment can be justified by data sensitivity, intranet requirements, complex permissions or multi-system integration. For a small document Q&A, SaaS or hybrid is usually faster and cheaper.
How do we reduce hallucinations in RAG?
Constrain answers to retrieved sources, require citations, say “unknown” when evidence is missing, and evaluate retrieval recall, faithfulness, citation accuracy and refusal with real questions. Prompts reduce risk but cannot replace permissions, data governance and human sampling.
How long does a knowledge base take to launch?
There is no universal timeline and we do not promise one. Scope, document quality, permissions and integration dominate; start with a controlled POC and expand only after acceptance. The hard part is document preparation, evaluation, access control and adoption — not installing a model.
Sources and caveats
Public prices and official documentation, accessed 25 October 2026. Prices can change by region, peak/off-peak, batch and promotion. The 300,000-question example and the RMB 220,000–650,000 first-year range are editorial assumptions, not a quote. Case-level accuracy and timelines depend on documents, permissions, models and evaluation; this article does not guarantee them.
- DeepSeek API Docs, Models & Pricing (accessed 25 Oct 2026): api-docs.deepseek.com. deepseek-flash cache-miss input RMB 1–2 per 1M tokens; output RMB 4–8 per 1M tokens (off-peak/peak).
- Alibaba Cloud Model Studio, text-embedding-v4 (updated 23 Jul 2026): help.aliyun.com. Beijing text input RMB 0.5 per 1M tokens; batch RMB 0.25 per 1M tokens.
- Alibaba Cloud Model Studio, Create a knowledge base (updated 10 Oct 2026): help.aliyun.com. Parsing, chunking, embedding and retrieval configuration; not a price list.
- Zilliz Cloud Pricing (accessed 25 Oct 2026): zilliz.com. Free tier includes 5 GB storage and 2.5M vCU/month; production is quoted by CU/storage/SLA.
- Huawei Cloud, RAG knowledge-assistant evaluation (updated 16 Sep 2026): support.huaweicloud.com. Hallucination, refusal and retrieval-noise evaluation approach.