Auto-Tag Documents with an LLM

This guide uses the retrieval and generation capabilities of the Knowledge Q&A service to have an LLM read every chunk of a document, extract tags, and write them back to that document. The result is a repeatable end-to-end pipeline: ingest, auto-tag, retrieve by tag. Every step runs in the console.

Overview

Why auto-tag

Manual tagging breaks down in three ways:

Tagging can't keep up with ingestion: When documents flow in as batches, human tagging lags far behind.
Documents sit unusable until tagged: Tag-based retrieval and routing wait on the human step, delaying time-to-value.
Tag definitions drift: Different annotators interpret categories differently, so similar documents end up with inconsistent tags and retrieval quality suffers.

Having an LLM extract tags directly from document content moves tagging into the ingestion pipeline and enforces a single tagging convention through the prompt.

When to use it

Scenario	Description	What auto-tagging adds
Enterprise document management	Tag contracts, project materials, and technical docs with department, project, topic, and confidentiality level	Speeds up search, supports access control and compliance audits
Knowledge asset curation	Tag technical docs, research reports, and industry white papers with domain, tech stack, and use case	Turns scattered material into a structured library, enabling precise tag-based Q&A and recommendation
Marketing asset management	Tag product manuals, marketing plans, and case studies with product line, industry, and scenario	Locate relevant material quickly, filter and reuse by tag

How it works

The Knowledge Q&A service is a RAG pipeline: retrieve chunks, then have the LLM answer based on those chunks. Swap the generation goal from "answer the question" to "extract tags" and you have auto-tagging. Give the target document a unique tag, use tag filtering to retrieve only chunks from that document, prompt the LLM to output tags, and write them back.

Three details make this work:

Single-document recall: Model Studio's tag filter matches on tags, not document ID. Adding a unique tag to the target document is what limits retrieval to that document.
Cover as much of the document as possible: Disable the rerank model and push recall count to the maximum so as many chunks as possible reach the LLM.
Structured output: The prompt constrains the model to emit parseable tags that can be written back directly.

Knowledge Q&A caps recall at 20 chunks per request. Documents with up to 20 chunks are covered in one pass; longer documents can't be read in a single request and must be split or tagged in batches. The small documents used in this walkthrough fit within the 20-chunk limit.

Walkthrough

This walkthrough uses the Model Studio docs sample bundle (92 official Model Studio technical documents). It tags First call to the Qwen API from the bundle, but any document of your own works too. The following tags are written by the LLM — no need to create them in advance:

Tag	Description	Example value
Product	The product this document covers	`product_model-studio`
Topic	Document topic	`topic_api-call`
Difficulty	Beginner / Intermediate / Expert	`difficulty_beginner`
Keyword	Document keywords (multi-value)	`keyword_environment-variable`, `keyword_curl`

Model Studio tags are a plain string array — there's no field-type concept. Tags may contain Chinese characters, English letters, digits, underscore _, and hyphen - only; colons, spaces, and other characters are not allowed. Each tag is capped at 32 characters, and each document can have at most 100 tags. Use the fieldname_value convention (for example product_model-studio) to express typed tags, and split multi-value dimensions like keywords into multiple strings.

1. Create a Knowledge Base and upload documents

Go to Data Ingestion → Knowledge Management and click Create. Choose the Flagship edition — auto-tagging needs a higher recall quota than the standard edition provides. Then walk through the three-step wizard:

Basic Information: Fill in the name and description. Set Knowledge Base type to Document Search and use case to Basic Document Q&A.
Select Data: Choose a connector (the default File connector is fine), select Upload Files as the data source, configure the category, and upload the sample documents in the File Upload area.
Index Settings: Keep the defaults and finish.

Open the new Knowledge Base's Document List and wait for each document's status to change to Parsed (parsing, chunking, and embedding run automatically).

The Add Tag field at the bottom of the Select Data step applies to every file in this upload — it can't isolate a single document. The unique per-document tag is added separately in the next step. Tags are free text and don't need to be created in advance.

Select Data step of the create-KB wizard: connector, data source, category, file upload

See Create a Knowledge Base and Document Management for details.

2. Add a unique tag to the target document

Tag filtering is how you limit retrieval to a single document, so the target document needs a unique tag (for example auto_tag_temp):

Console: In Data Ingestion → Knowledge Management, click View Details on the Knowledge Base you just created, open the document list, click More → Tags on the target document, enter auto_tag_temp in the tag dialog, press Enter, and confirm.
API: Call batchUpdateFileTag and set the document's tags to ["auto_tag_temp"].

3. Configure the Knowledge Q&A service

Go to Knowledge Service → Knowledge Q&A, click Create, and bind the Knowledge Base from the previous step.
Pick a generation model. Tag extraction demands strong comprehension and instruction-following, so choose a capable model such as qwen3.7-plus.
Click Expand Knowledge Base Configuration and set retrieval parameters as follows:

Parameter	Value	Why
Tag filter	Enter or pick the unique tag `auto_tag_temp` from the dropdown	Restricts recall to this document, enabling single-document tagging
Initial vector recall TopK	Increase to at least the document's chunk count (default 50)	Ensures the document's chunks are in the candidate pool
Initial keyword recall TopK	Increase to at least the document's chunk count (default 50)	Covers the keyword retrieval path as well
Rerank model	Select Do not use a model	Prevents rerank from filtering chunks by relevance
Max recall count	Push to the ceiling: 20	Sends as many chunks as possible to the model

Knowledge Base configuration: tag filter set to the unique tag, TopK increased, rerank disabled, max recall set to 20

Knowledge Base retrieval parameter configuration

The point of the tagging step is to let the model read as much of the document as possible. Set the rerank model to Do not use a model and push max recall to 20. Because of the 20-chunk cap, documents with more than 20 chunks need to be split or tagged in batches.

4. Write the tagging prompt

In the Prompt input on the configuration page, fill in a tagging prompt template. Retrieved chunks are stitched together into context and injected automatically by the service, so no placeholder is needed in the prompt itself. The prompt tells the model to emit a ready-to-write fieldname_value tag array:

# Task
You are a document-understanding assistant. The reference material you receive is a set of chunks from a single document, covering as much of that document as possible. Treat these chunks as different parts of one complete document and extract tags for it.

# Rules
1. Base your tags on the full set of chunks together, not on any single passage.
2. Every tag value must come from the reference material — do not use outside knowledge or subjective inference.
3. If a dimension is not clearly reflected in the material, omit that dimension.
4. Do not paraphrase the source, do not explain your reasoning — output only the tag array.

# Tag dimensions
- Product: the product this document covers (e.g. Model Studio, Qwen)
- Topic: document topic (e.g. api-call, model-inference, data-import)
- Difficulty: beginner / intermediate / expert
- Keyword: 1–5 keywords for the document

# Output format
- Represent each tag as fieldname_value. For multi-value dimensions (such as keyword), output multiple tags.
- Tags may contain Chinese characters, letters, digits, underscore, and hyphen only — no colons, spaces, or other characters.
- Output a single JSON string array, no Markdown, no comments, no extra text. For example:
["product_model-studio", "topic_api-call", "difficulty_beginner", "keyword_environment-variable", "keyword_curl"]

5. Run tagging in the console

In the Debug panel on the right of the configuration page, enter a tagging instruction (for example Extract tags for this document), send it, and check whether the tag array matches expectations.

Debug panel: send a tagging instruction, model returns a tag array

To tag many documents at once, confirm the results in the console, click Publish to get the service ID (agent_id), and drive tagging through the Knowledge Q&A API (the response is an SSE stream; see the API reference for parsing). Then write the tags back in the next step. For how to sequence tagging across documents in a batch, see Caveats.

6. Write tags back and retrieve by tag

Once you have the tag array from the model, call batchUpdateFileTag to write it back to the document. OVERWRITE replaces the temporary unique tag from Step 2:

curl -X POST 'https://{workspaceId}.cn-beijing.maas.aliyuncs.com/api/v1/connector/dash/batchUpdateFileTag' \
  -H 'Authorization: Bearer $API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
    "fileInfos": [
      {
        "fileId": "file-xxx",
        "tags": ["product_model-studio", "topic_api-call", "difficulty_beginner", "keyword_environment-variable", "keyword_curl"]
      }
    ],
    "updateMode": "OVERWRITE"
  }'

Get fileId from listFile, or update tags by hand in the console via More → Tags on the document row. After writing, set the corresponding tag in Knowledge Base Configuration → Tag Filter on the retrieval or Q&A service to filter recall by tag.

Document list shows the written tags; Knowledge Base configuration uses tag filter for precise recall

Caveats

Tags are a string array

Model Studio tags are a plain string array (no field-type concept). They may contain Chinese characters, letters, digits, underscore _, and hyphen - only — no colons, spaces, or other characters. Each tag is capped at 32 characters, and each document can have at most 100 tags. Use fieldname_value (for example product_model-studio) for typed tags, and split multi-value dimensions into multiple strings. Tags are free text and don't need to be created in advance.

Single-document recall depends on a unique tag

Model Studio retrieval and Q&A don't filter by document ID — tag filtering matches on tags. To tag a single document, add a unique tag (such as auto_tag_temp) to that document specifically. The Add Tag field in the Select Data step of the create-KB wizard applies to every file in the upload and can't isolate one document.

Bound by the max recall cap

Knowledge Q&A caps recall at 20 chunks per request — at most 20 chunks reach the model in one call. Documents with more than 20 chunks can't be covered in one pass and must be split or tagged in batches.

Tag filter is a service-level static config

The tag filter on a Q&A service is fixed once you publish and doesn't change per request. Two ways to tag many documents:

Approach	How	When to use
Shared temporary unique tag	Add the same temporary tag (such as `auto_tag_temp`) to every document, tag them one by one, then overwrite with the business tags	Small batches, sequential tagging
Multiple services in parallel	Create a separate Q&A service per document (or per group), each with its own tag filter	Large-scale parallel tagging

FAQ

Tagging results only cover part of the document? Check that rerank is set to Do not use a model, and that initial retrieval TopK and max recall are both increased. If the document has more than 20 chunks, split it or tag in batches. The model outputs extra text alongside the tag array? Tighten the "output only the array" constraint in the prompt. Before writing back, extract only the first [...] array from the output and parse that. Writing tags back returns an error? Check for unsupported characters (colons, spaces, etc.) and verify no tag exceeds 32 characters.

​Overview

​Why auto-tag

​When to use it

​How it works

​Walkthrough

​1. Create a Knowledge Base and upload documents

​2. Add a unique tag to the target document

​3. Configure the Knowledge Q&A service

​4. Write the tagging prompt

​5. Run tagging in the console

​6. Write tags back and retrieve by tag

​Caveats

​Tags are a string array

​Single-document recall depends on a unique tag

​Bound by the max recall cap

​Tag filter is a service-level static config

​FAQ

Overview

Why auto-tag

When to use it

How it works

Walkthrough

1. Create a Knowledge Base and upload documents

2. Add a unique tag to the target document

3. Configure the Knowledge Q&A service

4. Write the tagging prompt

5. Run tagging in the console

6. Write tags back and retrieve by tag

Caveats

Tags are a string array

Single-document recall depends on a unique tag

Bound by the max recall cap

Tag filter is a service-level static config

FAQ