Knowledge Base billing methods, cost breakdown, billing examples, and bill management
Alibaba Cloud Model Studio Knowledge Base service officially began billing on January 4, 2026. This document details the billing rules, including cost breakdown, calculation methods, and cost optimization strategies.
Knowledge Base data created before January 4, 2026 but not activated will be retained until June 30, 2026. After that date, the data will be permanently deleted. Please activate the Knowledge Base service in time.
Knowledge Bases support two billing methods: postpaid (pay-as-you-go) and prepaid resource packs. After activation, postpaid is the default. Resource packs can be purchased from the console: Standard resource pack or Enterprise resource pack.
Activation: Free.
Billing start: From the moment a Knowledge Base is successfully created.
Tier fees are runtime charges. Model Studio offers two Knowledge Base tiers: Standard and Enterprise.
If you use a self-purchased ADB-PG as vector storage, you will need to pay for it separately.
Standard: For personal / small-scale, PoC environments.
Enterprise: For high-concurrency, production-grade environments.
Tier
Max concurrency (retrieval)
Storage
Price
Standard
1 QPS (fixed, not adjustable)
Platform storage ≤ 100 GB
0.03 yuan / Knowledge Base / hour
Enterprise
50–10,000 QPS (adjustable, 1–200 RCU)
Platform storage ≤ 9,999 GB (for larger storage, configure a self-purchased ADB-PG instance at creation time)
0.2 yuan / RCU / hour
RCU: RCU (Retrieval Compute Unit) is the unit for retrieval concurrency. 1 RCU ≈ 50 QPS. More RCU means more concurrent capacity.
How to estimate RCU: Required RCU = ceil(peak QPS ÷ 50). For example, 80 QPS requires at least 2 RCU.
Max concurrency: Refers to the Knowledge Base's core retrieval performance (excluding dependent links, e.g. calling rerank models).
When retrieving from an Enterprise Knowledge Base, if model-dependent links are rate-limited under extreme conditions (e.g. Embedding, Rerank models), we will scale up the relevant services as quickly as possible, but a small number of retrieval requests may be degraded briefly.
Storage cost: The above prices include platform storage. If you use a self-purchased ADB-PG, you pay for it separately — see the ADB-PG pricing page.
Tier change: Billed in segments by the time of change. The change operation itself is free. Up to 1 tier change per Knowledge Base per calendar day.
Model Studio provides a one-time 720-hour free quota for all users. After the free quota is exhausted, pay-as-you-go applies.
Existing users: Free quota is valid until February 3, 2026 23:59, after which pay-as-you-go automatically applies.
New users: Free quota is valid for 30 days from activation. Unused quota expires after that.
Existing users are those who activated the service before January 4, 2026 (when billing officially began). New users are those who first activated on or after that date.
You can view remaining free hours and validity by clicking View Bill in the top-right corner of the Knowledge Base page.Usage rules
Scope: Only offsets Standard Knowledge Base tier fees — not Enterprise.
Deduction: Accumulated by actual runtime. Multiple Knowledge Bases consume quota proportionally.
For example, running 4 Standard Knowledge Bases simultaneously consumes 4 hours of quota per wall-clock hour.
Exclusions: Model call charges are not covered by the free quota — they follow each model's own billing policy.
Examples
Single Standard Knowledge Base: 720 hours ÷ 1 = 720 hours of free runtime
2 Standard Knowledge Bases: 720 hours ÷ 2 = 360 hours of free runtime
Multiple resource packs of the same type: The one expiring first is deducted first. If expiry dates are the same, the one purchased first is deducted first.
Overage: If all resource packs of a type expire or are exhausted, the excess automatically switches to pay-as-you-go.
Monitoring & alerts:
View remaining: Click Resource packs to view remaining amounts; click Statistics for usage details.
Set alerts: Budget alerts are recommended. When usage drops below a preset threshold, the system sends notifications via SMS, email, and in-site messages.
Refunds: Per Alibaba Cloud refund policy, unused portions of prepaid products can be refunded based on unused quota; used portions are non-refundable. After refund, the Knowledge Base switches to pay-as-you-go; deleting the Knowledge Base stops billing.
LLM for answer generation in the Q&A service, selected by the user in the app
Model call charges are a separate billing item, calculated by actual input Token volume. Prices and free quota policies follow the corresponding model's billing standard on the Model Marketplace — they are not included in Knowledge Base tier fees.
Billing formula: Model cost = (total input Tokens / 1000) × model unit price (yuan / 1K Tokens)Multiple Knowledge Bases: When an app has multiple Knowledge Bases mounted, retrieval is performed across all of them. Token consumption (Query embedding + Rerank) scales linearly with the number of Knowledge Bases (N Knowledge Bases = × N consumption).
Embedding: The embedding model vectorizes the user's Query.
Knowledge-Base routing (optional): If the app has multiple Knowledge Bases and routing is enabled, qwen-plus is called to determine which Knowledge Bases to route the query to — billed by qwen-plus Token usage.
Rerank (optional): The rerank model re-sorts preliminary results to improve accuracy. Document Knowledge Bases use qwen3-rerank; image and video Knowledge Bases use qwen3-vl-rerank.
Billing:
Query embedding cost: Based on the Token count of the user's input.
Rerank cost (can be disabled): This is the main retrieval cost. Cost depends on the total number of preliminary recalled chunks.
Retrieval flow and billing relationship
Preliminary recall: The system recalls text chunks based on:
Initial vector TopK: Controls the number of chunks recalled by semantic similarity (default 50)
Initial keyword TopK: Controls the number of chunks recalled by exact text match (default 50)
Reranking:
All preliminarily recalled chunks are sent to the Rerank model.
Cost = total preliminary chunks × average chunk Token count × model unit price
Rerank cost depends on the total number of preliminarily recalled chunks, not the number of finally returned chunks.
Final recall: After reranking, the system returns results based on the final recall count parameter (e.g. 5).
Clicking Hit Test on a Knowledge Base card to enter the debug page will incur model (embedding, rerank) call charges.
When using Knowledge Bases through Model Studio apps (agent apps, workflow apps) for Q&A, the following model call charges apply in addition to retrieval-stage fees:
Answer generation model: The system uses the Q&A model you select in the app (e.g. qwen-plus) to generate answers — billed by that model's Token usage.
In-chat file parsing (optional): When users upload files during a conversation and enable pre-file parsing, the system calls qwen3-rerank to process file content — billed by the rerank model's Token usage.
Knowledge-Base routing (optional): If the app has multiple Knowledge Bases and routing is enabled, qwen-plus is called for routing decisions.
The full cost of the Q&A service = tier fee (Knowledge Base runtime) + retrieval-stage model fees (embedding + rerank + routing) + Q&A-stage model fees (answer generation + in-chat file parsing). Each model fee is calculated independently by actual Token consumption.
The embedding models (e.g. text-embedding-v4) and rerank models (e.g. qwen3-rerank) used by Knowledge Bases belong to Model Studio's Class A models. Their call charges can be offset by the following savings plans:
AI General Savings Plan (recommended): Covers all Class A models (text embedding, multimodal embedding, rerank), with tiered discounts for monthly spending commitments.
Embedding & Rerank Savings Plan: A savings plan specifically for embedding and rerank models, purchased as a fixed amount upfront.
Savings plans only offset model call charges — they cannot offset Knowledge Base tier fees (runtime charges). For tier fee optimization, see Resource packs above.
The following examples are based on document search Knowledge Bases using text-embedding-v4 (embedding model) and qwen3-rerank (rerank model), both priced at 0.0005 yuan / 1K Tokens.Billing logic: Cost = Token consumption (in "1K Tokens") × model unit priceCreating a Knowledge Base
Action: Upload a file containing 50,000 Tokens for embedding.
Cost: 50 × 0.0005 yuan / 1K Tokens = 0.025 yuan
Updating a Knowledge Base
Action: Add a new file containing 20,000 Tokens.
Cost: 20 × 0.0005 yuan / 1K Tokens = 0.01 yuan
Retrieving from a single Knowledge Base
Action: Input a 100-Token Query; recall 150 chunks (avg 500 Tokens / chunk) for reranking.
Scenario: During 14:40–15:40, at 15:10 the tier is upgraded from Standard to Enterprise (2 RCU). Standard and Enterprise each run for 30 minutes (0.50 hours, rounded to 2 decimals).
Export from the Bill Details page. In the bill (aggregated by hour), view the tier fee for a specific Knowledge Base in the corresponding period (list total price column).
The instance ID in the bill is the Knowledge Base ID.
Filter by Product Name = Model Studio, Product = Knowledge Base (RAG) - Postpaid to view each Knowledge Base instance's billing item (e.g. Standard - Compute Resource) and corresponding list price (e.g. ¥0.03 yuan / (count × hour)).
Querying Token consumption and amounts in detail bills
Export from the Bill Details page. In the bill (aggregated by hour), view Token usage (usage column) and corresponding amount (list total price column) for the period.Viewing embedding model usageHover over the instance ID in the bill: if it looks like llm-xxx;xxx-embedding-xxx;embedding_token;RAG;0, it was generated by an embedding model.On the Bill Details page, set Bill Date to by month, filter Product Name = Model Studio, Product = Model Studio Model Inference, and Include zero-amount = Yes. The bill detail table shows Token consumption at hourly granularity, including service start/end time, billing item name, usage, unit (1K Tokens), list price, and list total price.Viewing rerank model usageHover over the instance ID: if it looks like llm-xxx;xxx-rerank;embedding_token;RAG;0, it was generated by a rerank model.On the Bill Details page, filter Product Name = Model Studio, Product = Model Studio Model Inference. In the bill detail table, the billing item name is "text vector usage" — view usage (unit: 1K tokens) and corresponding list total price for each period.
If you need to attribute costs to different departments or projects, you can use the "tag" feature to label workspaces.Step 1: Get workspace infoIn Workspace Management, identify the Workspace ID (e.g. llm-xxx) of the workspace to tag.Step 2: Bind tags
Select "Enter multiple resource IDs", search and select "Model Studio: Workspace" in the product tab, choose the workspace's region, enter the Workspace ID in the resource ID field, then click bind tags.
Create a tag key/value or use a preset tag to bind to the workspace, then click Confirm to complete.
After the operation, a Bind Resources confirmation popup will appear, showing each resource's ID, operation status, and failure reason. Verify and click Got it to close.
Step 3: VerifyYou have now completed binding tags to your Model Studio workspace. You can verify and query workspace tags via the Instance Tags column on the Bill Details page.
Newly created instance tags may have a delay (hourly level).When filtering, set Product Name = Model Studio, Product = Knowledge Base (RAG) - Postpaid.
After the Alibaba Cloud account becomes overdue, all Knowledge Bases enter a suspended state (cannot retrieve, update, or create via console or API) and stop billing.Platform storage
Phase
Description
0–14 days
Cannot retrieve, update, or create via console or API, but existing data is retained. Paying all overdue bills within the first 14 days will automatically restore normal service.
≥ 15 days
On the 15th day, the service is considered abandoned. Model Studio will release the Knowledge Bases and permanently delete their data — unrecoverable.
Self-purchased ADB-PG storage
Phase
Description
0–7 days
Cannot retrieve, update, or create via console or API, but existing data is retained. Paying all overdue bills within the first 7 days will automatically restore normal service.
≥ 8 days
On the 8th day, ADB-PG postpaid service is considered abandoned. ADB-PG will clean up Knowledge Base instances and permanently delete their data — unrecoverable.
When using self-purchased ADB-PG, the data retention period follows ADB-PG's overdue policy: 7 days (not 14).
Can a sub-account activate a Knowledge Base or view bills?
Yes. An authorized sub-account (system policy AliyunBailianFullAccess or AliyunSFMFullAccess) can activate Knowledge Bases. Fees are charged to the primary account.
What if the Enterprise tier's platform storage is not enough for my data?
You can choose to use a self-purchased ADB-PG instance as vector storage when creating a Knowledge Base. See Create a Knowledge Base for configuration details.
How is billing handled when a tier change spans an hour boundary?
Billing is segmented by the time of change — within the same hour, each time segment is proportionally accumulated. See Tier change (segmented billing) above for an example.
Why are my rerank charges so high? How can I reduce model call costs?
Rerank model charges are determined by the total number of preliminarily recalled chunks, not the number of results returned. See Cost optimization tips above for ways to reduce model call costs.
How do I completely stop Knowledge Base billing? Can I just delete files?
No. The only way to stop billing is to delete the entire Knowledge Base instance.
Wrong approach: Deleting only the files within a Knowledge Base just clears the data — the Knowledge Base instance (as the billing entity) is still running, so tier fees continue to accrue.
Correct approach: Find the Knowledge Base instance in the console and delete it.
Deleting a Knowledge Base permanently removes all data and cannot be undone. Proceed with caution.
Why are rerank model call counts higher than app call counts?
This is an automatic system optimization for performance. When a single request to the rerank model contains a large number of chunks, the system splits it into multiple batches to speed up processing.This increases the rerank model's call count, but the total cost remains the same — billing is based on total Token consumption, not call count.