feat: Adding support for customizing chunk context in RAG insertion and querying (#2134)

# What does this PR do?
his PR allows users to customize the template used for chunks when
inserted into the context. Additionally, this enables metadata injection
into the context of an LLM for RAG. This makes a naive and crude
assumption that each chunk should include the metadata, this is
obviously redundant when multiple chunks are returned from the same
document. In order to remove any sort of duplication of chunks, we'd
have to make much more significant changes so this is a reasonable first
step that unblocks users requesting this enhancement in
https://github.com/meta-llama/llama-stack/issues/1767.

In the future, this can be extended to support citations.


List of Changes:
- `llama_stack/apis/tools/rag_tool.py`
    - Added  `chunk_template` field in `RAGQueryConfig`.
- Added `field_validator` to validate the `chunk_template` field in
`RAGQueryConfig`.
- Ensured the `chunk_template` field includes placeholders `{index}` and
`{chunk.content}`.
- Updated the `query` method to use the `chunk_template` for formatting
chunk text content.
- `llama_stack/providers/inline/tool_runtime/rag/memory.py`
- Modified the `insert` method to pass `doc.metadata` for chunk
creation.
- Enhanced the `query` method to format results using `chunk_template`
and exclude unnecessary metadata fields like `token_count`.
- `llama_stack/providers/utils/memory/vector_store.py`
- Updated `make_overlapped_chunks` to include metadata serialization and
token count for both content and metadata.
    - Added error handling for metadata serialization issues.
- `pyproject.toml`
- Added `pydantic.field_validator` as a recognized `classmethod`
decorator in the linting configuration.
- `tests/integration/tool_runtime/test_rag_tool.py`
- Refactored test assertions to separate `assert_valid_chunk_response`
and `assert_valid_text_response`.
- Added integration tests to validate `chunk_template` functionality
with and without metadata inclusion.
- Included a test case to ensure `chunk_template` validation errors are
raised appropriately.
- `tests/unit/rag/test_vector_store.py`
- Added unit tests for `make_overlapped_chunks`, verifying chunk
creation with overlapping tokens and metadata integrity.
- Added tests to handle metadata serialization errors, ensuring proper
exception handling.
- `docs/_static/llama-stack-spec.html`
- Added a new `chunk_template` field of type `string` with a default
template for formatting retrieved chunks in RAGQueryConfig.
    - Updated the `required` fields to include `chunk_template`.
- `docs/_static/llama-stack-spec.yaml`
- Introduced `chunk_template` field with a default value for
RAGQueryConfig.
- Updated the required configuration list to include `chunk_template`.
- `docs/source/building_applications/rag.md`
- Documented the `chunk_template` configuration, explaining how to
customize metadata formatting in RAG queries.
- Added examples demonstrating the usage of the `chunk_template` field
in RAG tool queries.
    - Highlighted default values for `RAG` agent configurations.

# Resolves https://github.com/meta-llama/llama-stack/issues/1767

## Test Plan
Updated both `test_vector_store.py` and `test_rag_tool.py` and tested
end-to-end with a script.

I also tested the quickstart to enable this and specified this metadata:
```python
document = RAGDocument(
    document_id="document_1",
    content=source,
    mime_type="text/html",
    metadata={"author": "Paul Graham", "title": "How to do great work"},
)
```
Which produced the output below: 

![Screenshot 2025-05-13 at 10 53
43 PM](https://github.com/user-attachments/assets/bb199d04-501e-4217-9c44-4699d43d5519)

This highlights the usefulness of the additional metadata. Notice how
the metadata is redundant for different chunks of the same document. I
think we can update that in a subsequent PR.

# Documentation
I've added a brief comment about this in the documentation to outline
this to users and updated the API documentation.

---------

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
This commit is contained in:
Francisco Arceo 2025-05-14 19:56:20 -06:00 committed by GitHub
parent ff247e35be
commit 8e7ab146f8
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 230 additions and 29 deletions

View file

@ -11294,24 +11294,34 @@
"type": "object",
"properties": {
"query_generator_config": {
"$ref": "#/components/schemas/RAGQueryGeneratorConfig"
"$ref": "#/components/schemas/RAGQueryGeneratorConfig",
"description": "Configuration for the query generator."
},
"max_tokens_in_context": {
"type": "integer",
"default": 4096
"default": 4096,
"description": "Maximum number of tokens in the context."
},
"max_chunks": {
"type": "integer",
"default": 5
"default": 5,
"description": "Maximum number of chunks to retrieve."
},
"chunk_template": {
"type": "string",
"default": "Result {index}\nContent: {chunk.content}\nMetadata: {metadata}\n",
"description": "Template for formatting each retrieved chunk in the context. Available placeholders: {index} (1-based chunk ordinal), {chunk.content} (chunk content string), {metadata} (chunk metadata dict). Default: \"Result {index}\\nContent: {chunk.content}\\nMetadata: {metadata}\\n\""
}
},
"additionalProperties": false,
"required": [
"query_generator_config",
"max_tokens_in_context",
"max_chunks"
"max_chunks",
"chunk_template"
],
"title": "RAGQueryConfig"
"title": "RAGQueryConfig",
"description": "Configuration for the RAG query generation."
},
"RAGQueryGeneratorConfig": {
"oneOf": [

View file

@ -7794,18 +7794,37 @@ components:
properties:
query_generator_config:
$ref: '#/components/schemas/RAGQueryGeneratorConfig'
description: Configuration for the query generator.
max_tokens_in_context:
type: integer
default: 4096
description: Maximum number of tokens in the context.
max_chunks:
type: integer
default: 5
description: Maximum number of chunks to retrieve.
chunk_template:
type: string
default: >
Result {index}
Content: {chunk.content}
Metadata: {metadata}
description: >-
Template for formatting each retrieved chunk in the context. Available
placeholders: {index} (1-based chunk ordinal), {chunk.content} (chunk
content string), {metadata} (chunk metadata dict). Default: "Result {index}\nContent:
{chunk.content}\nMetadata: {metadata}\n"
additionalProperties: false
required:
- query_generator_config
- max_tokens_in_context
- max_chunks
- chunk_template
title: RAGQueryConfig
description: >-
Configuration for the RAG query generation.
RAGQueryGeneratorConfig:
oneOf:
- $ref: '#/components/schemas/DefaultRAGQueryGeneratorConfig'

View file

@ -51,6 +51,7 @@ chunks = [
"mime_type": "text/plain",
"metadata": {
"document_id": "doc1",
"author": "Jane Doe",
},
},
]
@ -98,6 +99,17 @@ results = client.tool_runtime.rag_tool.query(
)
```
You can configure how the RAG tool adds metadata to the context if you find it useful for your application. Simply add:
```python
# Query documents
results = client.tool_runtime.rag_tool.query(
vector_db_ids=[vector_db_id],
content="What do you know about...",
query_config={
"chunk_template": "Result {index}\nContent: {chunk.content}\nMetadata: {metadata}\n",
},
)
```
### Building RAG-Enhanced Agents
One of the most powerful patterns is combining agents with RAG capabilities. Here's a complete example:
@ -115,6 +127,12 @@ agent = Agent(
"name": "builtin::rag/knowledge_search",
"args": {
"vector_db_ids": [vector_db_id],
# Defaults
"query_config": {
"chunk_size_in_tokens": 512,
"chunk_overlap_in_tokens": 0,
"chunk_template": "Result {index}\nContent: {chunk.content}\nMetadata: {metadata}\n",
},
},
}
],

View file

@ -7,7 +7,7 @@
from enum import Enum
from typing import Annotated, Any, Literal
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, field_validator
from typing_extensions import Protocol, runtime_checkable
from llama_stack.apis.common.content_types import URL, InterleavedContent
@ -67,11 +67,33 @@ register_schema(RAGQueryGeneratorConfig, name="RAGQueryGeneratorConfig")
@json_schema_type
class RAGQueryConfig(BaseModel):
"""
Configuration for the RAG query generation.
:param query_generator_config: Configuration for the query generator.
:param max_tokens_in_context: Maximum number of tokens in the context.
:param max_chunks: Maximum number of chunks to retrieve.
:param chunk_template: Template for formatting each retrieved chunk in the context.
Available placeholders: {index} (1-based chunk ordinal), {chunk.content} (chunk content string), {metadata} (chunk metadata dict).
Default: "Result {index}\\nContent: {chunk.content}\\nMetadata: {metadata}\\n"
"""
# This config defines how a query is generated using the messages
# for memory bank retrieval.
query_generator_config: RAGQueryGeneratorConfig = Field(default=DefaultRAGQueryGeneratorConfig())
max_tokens_in_context: int = 4096
max_chunks: int = 5
chunk_template: str = "Result {index}\nContent: {chunk.content}\nMetadata: {metadata}\n"
@field_validator("chunk_template")
def validate_chunk_template(cls, v: str) -> str:
if "{chunk.content}" not in v:
raise ValueError("chunk_template must contain {chunk.content}")
if "{index}" not in v:
raise ValueError("chunk_template must contain {index}")
if len(v) == 0:
raise ValueError("chunk_template must not be empty")
return v
@runtime_checkable

View file

@ -87,6 +87,7 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
content,
chunk_size_in_tokens,
chunk_size_in_tokens // 4,
doc.metadata,
)
)
@ -142,19 +143,21 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
text=f"knowledge_search tool found {len(chunks)} chunks:\nBEGIN of knowledge_search tool results.\n"
)
]
for i, c in enumerate(chunks):
metadata = c.metadata
for i, chunk in enumerate(chunks):
metadata = chunk.metadata
tokens += metadata["token_count"]
tokens += metadata["metadata_token_count"]
if tokens > query_config.max_tokens_in_context:
log.error(
f"Using {len(picked)} chunks; reached max tokens in context: {tokens}",
)
break
picked.append(
TextContentItem(
text=f"Result {i + 1}:\nDocument_id:{metadata['document_id'][:5]}\nContent: {c.content}\n",
)
)
metadata_subset = {k: v for k, v in metadata.items() if k not in ["token_count", "metadata_token_count"]}
text_content = query_config.chunk_template.format(index=i + 1, chunk=chunk, metadata=metadata_subset)
picked.append(TextContentItem(text=text_content))
picked.append(TextContentItem(text="END of knowledge_search tool results.\n"))
picked.append(
TextContentItem(

View file

@ -139,22 +139,32 @@ async def content_from_doc(doc: RAGDocument) -> str:
return interleaved_content_as_str(doc.content)
def make_overlapped_chunks(document_id: str, text: str, window_len: int, overlap_len: int) -> list[Chunk]:
def make_overlapped_chunks(
document_id: str, text: str, window_len: int, overlap_len: int, metadata: dict[str, Any]
) -> list[Chunk]:
tokenizer = Tokenizer.get_instance()
tokens = tokenizer.encode(text, bos=False, eos=False)
try:
metadata_string = str(metadata)
except Exception as e:
raise ValueError("Failed to serialize metadata to string") from e
metadata_tokens = tokenizer.encode(metadata_string, bos=False, eos=False)
chunks = []
for i in range(0, len(tokens), window_len - overlap_len):
toks = tokens[i : i + window_len]
chunk = tokenizer.decode(toks)
chunk_metadata = metadata.copy()
chunk_metadata["document_id"] = document_id
chunk_metadata["token_count"] = len(toks)
chunk_metadata["metadata_token_count"] = len(metadata_tokens)
# chunk is a string
chunks.append(
Chunk(
content=chunk,
metadata={
"token_count": len(toks),
"document_id": document_id,
},
metadata=chunk_metadata,
)
)

View file

@ -320,3 +320,6 @@ ignore_missing_imports = true
init_forbid_extra = true
init_typed = true
warn_required_dynamic_aliases = true
[tool.ruff.lint.pep8-naming]
classmethod-decorators = ["classmethod", "pydantic.field_validator"]

View file

@ -49,7 +49,7 @@ def sample_documents():
]
def assert_valid_response(response):
def assert_valid_chunk_response(response):
assert len(response.chunks) > 0
assert len(response.scores) > 0
assert len(response.chunks) == len(response.scores)
@ -57,6 +57,11 @@ def assert_valid_response(response):
assert isinstance(chunk.content, str)
def assert_valid_text_response(response):
assert len(response.content) > 0
assert all(isinstance(chunk.text, str) for chunk in response.content)
def test_vector_db_insert_inline_and_query(client_with_empty_registry, sample_documents, embedding_model_id):
vector_db_id = "test_vector_db"
client_with_empty_registry.vector_dbs.register(
@ -77,7 +82,7 @@ def test_vector_db_insert_inline_and_query(client_with_empty_registry, sample_do
vector_db_id=vector_db_id,
query=query1,
)
assert_valid_response(response1)
assert_valid_chunk_response(response1)
assert any("Python" in chunk.content for chunk in response1.chunks)
# Query with semantic similarity
@ -86,7 +91,7 @@ def test_vector_db_insert_inline_and_query(client_with_empty_registry, sample_do
vector_db_id=vector_db_id,
query=query2,
)
assert_valid_response(response2)
assert_valid_chunk_response(response2)
assert any("neural networks" in chunk.content.lower() for chunk in response2.chunks)
# Query with limit on number of results (max_chunks=2)
@ -96,7 +101,7 @@ def test_vector_db_insert_inline_and_query(client_with_empty_registry, sample_do
query=query3,
params={"max_chunks": 2},
)
assert_valid_response(response3)
assert_valid_chunk_response(response3)
assert len(response3.chunks) <= 2
# Query with threshold on similarity score
@ -106,7 +111,7 @@ def test_vector_db_insert_inline_and_query(client_with_empty_registry, sample_do
query=query4,
params={"score_threshold": 0.01},
)
assert_valid_response(response4)
assert_valid_chunk_response(response4)
assert all(score >= 0.01 for score in response4.scores)
@ -126,9 +131,6 @@ def test_vector_db_insert_from_url_and_query(client_with_empty_registry, sample_
available_vector_dbs = [vector_db.identifier for vector_db in client_with_empty_registry.vector_dbs.list()]
assert vector_db_id in available_vector_dbs
# URLs of documents to insert
# TODO: Move to test/memory/resources then update the url to
# https://raw.githubusercontent.com/meta-llama/llama-stack/main/tests/memory/resources/{url}
urls = [
"memory_optimizations.rst",
"chat.rst",
@ -155,7 +157,7 @@ def test_vector_db_insert_from_url_and_query(client_with_empty_registry, sample_
vector_db_id=vector_db_id,
query="What's the name of the fine-tunning method used?",
)
assert_valid_response(response1)
assert_valid_chunk_response(response1)
assert any("lora" in chunk.content.lower() for chunk in response1.chunks)
# Query for the name of model
@ -163,5 +165,69 @@ def test_vector_db_insert_from_url_and_query(client_with_empty_registry, sample_
vector_db_id=vector_db_id,
query="Which Llama model is mentioned?",
)
assert_valid_response(response2)
assert_valid_chunk_response(response2)
assert any("llama2" in chunk.content.lower() for chunk in response2.chunks)
def test_rag_tool_insert_and_query(client_with_empty_registry, embedding_model_id):
providers = [p for p in client_with_empty_registry.providers.list() if p.api == "vector_io"]
assert len(providers) > 0
vector_db_id = "test_vector_db"
client_with_empty_registry.vector_dbs.register(
vector_db_id=vector_db_id,
embedding_model=embedding_model_id,
embedding_dimension=384,
)
available_vector_dbs = [vector_db.identifier for vector_db in client_with_empty_registry.vector_dbs.list()]
assert vector_db_id in available_vector_dbs
urls = [
"memory_optimizations.rst",
"chat.rst",
"llama3.rst",
]
documents = [
Document(
document_id=f"num-{i}",
content=f"https://raw.githubusercontent.com/pytorch/torchtune/main/docs/source/tutorials/{url}",
mime_type="text/plain",
metadata={"author": "llama", "source": url},
)
for i, url in enumerate(urls)
]
client_with_empty_registry.tool_runtime.rag_tool.insert(
documents=documents,
vector_db_id=vector_db_id,
chunk_size_in_tokens=512,
)
response_with_metadata = client_with_empty_registry.tool_runtime.rag_tool.query(
vector_db_ids=[vector_db_id],
content="What is the name of the method used for fine-tuning?",
)
assert_valid_text_response(response_with_metadata)
assert any("metadata:" in chunk.text.lower() for chunk in response_with_metadata.content)
response_without_metadata = client_with_empty_registry.tool_runtime.rag_tool.query(
vector_db_ids=[vector_db_id],
content="What is the name of the method used for fine-tuning?",
query_config={
"include_metadata_in_content": True,
"chunk_template": "Result {index}\nContent: {chunk.content}\n",
},
)
assert_valid_text_response(response_without_metadata)
assert not any("metadata:" in chunk.text.lower() for chunk in response_without_metadata.content)
with pytest.raises(ValueError):
client_with_empty_registry.tool_runtime.rag_tool.query(
vector_db_ids=[vector_db_id],
content="What is the name of the method used for fine-tuning?",
query_config={
"chunk_template": "This should raise a ValueError because it is missing the proper template variables",
},
)

View file

@ -12,7 +12,7 @@ from pathlib import Path
import pytest
from llama_stack.apis.tools import RAGDocument
from llama_stack.providers.utils.memory.vector_store import URL, content_from_doc
from llama_stack.providers.utils.memory.vector_store import URL, content_from_doc, make_overlapped_chunks
DUMMY_PDF_PATH = Path(os.path.abspath(__file__)).parent / "fixtures" / "dummy.pdf"
# Depending on the machine, this can get parsed a couple of ways
@ -76,3 +76,53 @@ class TestVectorStore:
)
content = await content_from_doc(doc)
assert content in DUMMY_PDF_TEXT_CHOICES
@pytest.mark.parametrize(
"window_len, overlap_len, expected_chunks",
[
(5, 2, 4), # Create 4 chunks with window of 5 and overlap of 2
(4, 1, 4), # Create 4 chunks with window of 4 and overlap of 1
],
)
def test_make_overlapped_chunks(self, window_len, overlap_len, expected_chunks):
document_id = "test_doc_123"
text = "This is a sample document for testing the chunking behavior"
original_metadata = {"source": "test", "date": "2023-01-01", "author": "llama"}
len_metadata_tokens = 24 # specific to the metadata above
chunks = make_overlapped_chunks(document_id, text, window_len, overlap_len, original_metadata)
assert len(chunks) == expected_chunks
# Check that each chunk has the right metadata
for chunk in chunks:
# Original metadata should be preserved
assert chunk.metadata["source"] == "test"
assert chunk.metadata["date"] == "2023-01-01"
assert chunk.metadata["author"] == "llama"
# New metadata should be added
assert chunk.metadata["document_id"] == document_id
assert "token_count" in chunk.metadata
assert isinstance(chunk.metadata["token_count"], int)
assert chunk.metadata["token_count"] > 0
assert chunk.metadata["metadata_token_count"] == len_metadata_tokens
def test_raise_overlapped_chunks_metadata_serialization_error(self):
document_id = "test_doc_ex"
text = "Some text"
window_len = 5
overlap_len = 2
class BadMetadata:
def __repr__(self):
raise TypeError("Cannot convert to string")
problematic_metadata = {"bad_metadata_example": BadMetadata()}
with pytest.raises(ValueError) as excinfo:
make_overlapped_chunks(document_id, text, window_len, overlap_len, problematic_metadata)
assert str(excinfo.value) == "Failed to serialize metadata to string"
assert isinstance(excinfo.value.__cause__, TypeError)
assert str(excinfo.value.__cause__) == "Cannot convert to string"