RAG Agent billing-ApsaraDB RDS(RDS)-阿里云帮助中心

RAG Agent charges are based on the number and type of API calls. This page explains the free quota, pricing tiers, API-to-call-count mapping, and a worked example so you can estimate costs before you start.

Free quota

Each day, the first 1,000 API calls are free. The free quota resets daily. Calls beyond 1,000 in a single day are billed at the rates listed below.

Billing method

RAG Agent uses pay-as-you-go billing. Charges are calculated as:

Total cost = Σ (Unit price of each API billing type × Number of calls)

Charges appear on your bill under the billable item RDS-AI Application Data Service.

API billing types and pricing

APIs are grouped into four billing types based on the underlying AI models they use. Each type has a fixed unit price per call.

API billing type	Primary capabilities	Underlying models	Unit price (CNY/call)
Intermediate	Embedding, Rerank	`qwen3-embedding-v4`, `gte-rerank-v2`	0.0006
Advanced	Knowledge graph construction (basic Large Language Models (LLMs)), Retrieved context summarization, Text parsing	`qwen3-flash`, `qwen3-vl-flash`	0.0013
Expert	Knowledge graph construction (basic LLMs), Retrieved context summarization	`qwen3-qwq`	0.0096
Master	Multimodal parsing	`qwen3-vl-plus`	0.0210

RAG capabilities and API mapping

Use this table to understand which capabilities each API invokes and roughly how many calls to expect per operation.

Call count key:

Symbol	Meaning
`-`	Not applicable
`✓`	1 call
`✓✓`	A few calls, typically 2–5, depending on the number of datasets
`✓✓✓`	Multiple calls, typically 5–10, depending on the number of document chunks
`✓✓✓✓`	Many calls, typically 10–100+, depending on the number of chunks, entities, and relations

API	Text parsing	Multimodal parsing	Embedding	Knowledge graph construction	Retrieved context summarization	Rerank
`documents/upload` (Standard)	-	-	✓✓✓	✓✓✓✓	-	-
`documents/upload` (Multimodal)	✓	✓✓	✓✓✓	✓✓✓✓	-	-
`documents/upload` (Audio/Video)	-	-	✓✓✓	✓✓✓✓	-	-
`documents/text`	-	-	✓✓✓	✓✓✓✓	-	-
`query`	-	-	✓	-	✓✓	✓
`query/stream`	-	-	✓	-	✓✓	✓
`cross-query`	-	-	✓✓	-	✓✓	✓✓
`cross-query/context`	-	-	✓✓	-	✓	✓✓
`cross-query/stream`	-	-	✓✓	-	✓✓	✓✓
`search`	-	-	✓	-	-	-

For document upload operations, knowledge graph construction dominates the total call count and cost. It can generate 10–100+ calls per upload depending on the number of chunks, entities, and relations extracted.

Billing example

This example estimates the cost of uploading a PDF document with mixed text and images, then running one retrieval query in mix mode with reranking enabled.

Scenario parameters

Parameter	Value	Description
Document word count	40,000 Chinese characters	Approximately 53,000 tokens (1 Chinese character ≈ 1.3 tokens)
Number of images	8	-
`chunk_token_size`	1200	Maximum tokens per chunk
`chunk_overlap_token_size`	100	Overlapping tokens between chunks
`entity_extract_max_gleaning`	1	Supplementary extraction rounds for the knowledge graph
Estimated entities	80	Estimated based on document content and model capabilities
Estimated relations	60	Estimated based on document content and model capabilities

Step 1: Calculate the number of chunks

The system splits the document into chunks before processing.

Effective chunk size: chunk_token_size − chunk_overlap_token_size = 1200 − 100 = 1,100 tokens
Estimated chunks: 53,000 ÷ 1,100 ≈ 48.18, rounded up to 49 chunks

Step 2: Upload phase — API calls and cost

Uploading this document via documents/upload (Multimodal) triggers text parsing, multimodal parsing, embedding, and knowledge graph construction.

Call count breakdown:

Text parsing: 1 call (parses the PDF structure)
Multimodal parsing: 8 calls (1 per image)
Embedding: 49 chunks + 80 entities + 60 relations = 189 calls
Knowledge graph construction: 49 chunks + 80 entities + 60 relations = 189 calls

Cost summary:

API billing type	Included capabilities	Call count	Unit price (CNY/call)	Cost (CNY)
Intermediate	Embedding (chunks, entities, relations)	189	0.0006	0.1134
Advanced	Knowledge graph construction, Text parsing	190	0.0013	0.2470
Master	Multimodal parsing	8	0.0210	0.1680
Total for upload phase		387	-	0.5284

Step 3: Query phase — API calls and cost

One query in mix mode with reranking enabled triggers embedding, reranking, and retrieved context summarization. In mix mode, the system retrieves from multiple knowledge sources (for example, text chunks and entities), which increases embedding calls.