# 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:
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>
# What does this PR do?
Checks for RAGDocument of type InterleavedContent
I noticed when stepping through the code that the supported types for
`RAGDocument` included `InterleavedContent` as a content type. This type
is not checked against before putting the `doc.content` is regex matched
against. This would cause a runtime error. This change adds an explicit
check for type.
The only other part that I'm unclear on is how to handle the
`ImageContent` type since this would always just return `<image>` which
seems like an undesired behavior. Should the `InterleavedContent` type
be removed from `RAGDocument` and replaced with `URI | str`?
## Test Plan
[//]: # (## Documentation)
---------
Signed-off-by: Kevin <kpostlet@redhat.com>
# What does this PR do?
The goal of this PR is code base modernization.
Schema reflection code needed a minor adjustment to handle UnionTypes
and collections.abc.AsyncIterator. (Both are preferred for latest Python
releases.)
Note to reviewers: almost all changes here are automatically generated
by pyupgrade. Some additional unused imports were cleaned up. The only
change worth of note can be found under `docs/openapi_generator` and
`llama_stack/strong_typing/schema.py` where reflection code was updated
to deal with "newer" types.
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
Fix import errors due to `chardet` and `pypdf` not being installed while
imported from `url_utils.py`.
Closes#1432
## Test Plan
Now able to run the server with the config.
[//]: # (## Documentation)
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
- Configured ruff linter to automatically fix import sorting issues.
- Set --exit-non-zero-on-fix to ensure non-zero exit code when fixes are
applied.
- Enabled the 'I' selection to focus on import-related linting rules.
- Ran the linter, and formatted all codebase imports accordingly.
- Removed the black dep from the "dev" group since we use ruff
Signed-off-by: Sébastien Han <seb@redhat.com>
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
[Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.*]
[//]: # (## Documentation)
[//]: # (- [ ] Added a Changelog entry if the change is significant)
Signed-off-by: Sébastien Han <seb@redhat.com>
Lint check in main branch is failing. This fixes the lint check after we
moved to ruff in https://github.com/meta-llama/llama-stack/pull/921. We
need to move to a `ruff.toml` file as well as fixing and ignoring some
additional checks.
Signed-off-by: Yuan Tang <terrytangyuan@gmail.com>
See https://github.com/meta-llama/llama-stack/issues/827 for the broader
design.
Third part:
- we need to make `tool_runtime.rag_tool.query_context()` and
`tool_runtime.rag_tool.insert_documents()` methods work smoothly with
complete type safety. To that end, we introduce a sub-resource path
`tool-runtime/rag-tool/` and make changes to the resolver to make things
work.
- the PR updates the agents implementation to directly call these typed
APIs for memory accesses rather than going through the complex, untyped
"invoke_tool" API. the code looks much nicer and simpler (expectedly.)
- there are a number of hacks in the server resolver implementation
still, we will live with some and fix some
Note that we must make sure the client SDKs are able to handle this
subresource complexity also. Stainless has support for subresources, so
this should be possible but beware.
## Test Plan
Our RAG test is sad (doesn't actually test for actual RAG output) but I
verified that the implementation works. I will work on fixing the RAG
test afterwards.
```bash
pytest -s -v tests/agents/test_agents.py -k "rag and together" --safety-shield=meta-llama/Llama-Guard-3-8B
```
See https://github.com/meta-llama/llama-stack/issues/827 for the broader
design.
Second part:
- updates routing table / router code
- updates the faiss implementation
## Test Plan
```
pytest -s -v -k sentence test_vector_io.py --env EMBEDDING_DIMENSION=384
```
## What does this PR do?
This is a long-pending change and particularly important to get done
now.
Specifically:
- we cannot "localize" (aka download) any URLs from media attachments
anywhere near our modeling code. it must be done within llama-stack.
- `PIL.Image` is infesting all our APIs via `ImageMedia ->
InterleavedTextMedia` and that cannot be right at all. Anything in the
API surface must be "naturally serializable". We need a standard `{
type: "image", image_url: "<...>" }` which is more extensible
- `UserMessage`, `SystemMessage`, etc. are moved completely to
llama-stack from the llama-models repository.
See https://github.com/meta-llama/llama-models/pull/244 for the
corresponding PR in llama-models.
## Test Plan
```bash
cd llama_stack/providers/tests
pytest -s -v -k "fireworks or ollama or together" inference/test_vision_inference.py
pytest -s -v -k "(fireworks or ollama or together) and llama_3b" inference/test_text_inference.py
pytest -s -v -k chroma memory/test_memory.py \
--env EMBEDDING_DIMENSION=384 --env CHROMA_DB_PATH=/tmp/foobar
pytest -s -v -k fireworks agents/test_agents.py \
--safety-shield=meta-llama/Llama-Guard-3-8B \
--inference-model=meta-llama/Llama-3.1-8B-Instruct
```
Updated the client sdk (see PR ...), installed the SDK in the same
environment and then ran the SDK tests:
```bash
cd tests/client-sdk
LLAMA_STACK_CONFIG=together pytest -s -v agents/test_agents.py
LLAMA_STACK_CONFIG=ollama pytest -s -v memory/test_memory.py
# this one needed a bit of hacking in the run.yaml to ensure I could register the vision model correctly
INFERENCE_MODEL=llama3.2-vision:latest LLAMA_STACK_CONFIG=ollama pytest -s -v inference/test_inference.py
```
This PR does the following:
1) adds the ability to generate embeddings in all supported inference
providers.
2) Moves all the memory providers to use the inference API and improved
the memory tests to setup the inference stack correctly and use the
embedding models
This is a merge from #589 and #598
# What does this PR do?
Addresses issue (#342)
- PDFs uploaded from url are being loaded into vector db as raw bytes
- Instead this PR extracts text from PDF if mime_type is
"application/json"
- Adds tests to cover new cases
## Test Plan
Ran these unit tests:
```bash
llama stack build --template meta-reference-gpu --image-type conda
conda activate llamastack-meta-reference-gpu
pip install pytest pytest-asyncio pypdf
pytest llama_stack/providers/tests/memory/test_vector_store.py -v
```
```
platform linux -- Python 3.10.15, pytest-8.3.3, pluggy-1.5.0 -- /home/ubuntu/1xa100-2/llama-stack/envs/bin/python
cachedir: .pytest_cache
rootdir: /home/ubuntu/1xa100-2/llama-stack
configfile: pyproject.toml
plugins: anyio-4.6.2.post1, asyncio-0.24.0, httpx-0.35.0
asyncio: mode=strict, default_loop_scope=None
collected 3 items
llama_stack/providers/tests/memory/test_vector_store.py::TestVectorStore::test_returns_content_from_pdf_data_uri PASSED [ 33%]
llama_stack/providers/tests/memory/test_vector_store.py::TestVectorStore::test_downloads_pdf_and_returns_content PASSED [ 66%]
llama_stack/providers/tests/memory/test_vector_store.py::TestVectorStore::test_downloads_pdf_and_returns_content_with_url_object PASSED [100%]
======================================================= 3 passed, 1 warning in 0.62s =======================================================
```
Tested manually via [this
script](afc8f8bebf/init.py)
to initialize and [this
script](afc8f8bebf/query.py)
to query
```bash
# Ran with meta-reference-gpu with safety
llama stack build --template meta-reference-gpu --image-type conda && llama stack run distributions/meta-reference-gpu/run-with-safety.yaml \
--port 5001 \
--env INFERENCE_MODEL=meta-llama/Llama-3.2-11B-Vision-Instruct
# Run init.py script
wget https://raw.githubusercontent.com/aidando73/llama-stack/afc8f8bebf70e1ad065d87e84692e1a3a45d9e19/init.py
pip install httpx==0.27.2 # Due to issue https://github.com/meta-llama/llama-stack-client-python/issues/54
python init.py
# Run query.py script
wget https://raw.githubusercontent.com/aidando73/llama-stack/afc8f8bebf70e1ad065d87e84692e1a3a45d9e19/query.py
python query.py
```
Should output valid text chunks
```
Chunk(content=' that it has a significantly\nlower violation rate than the competing standalone open source model, trading off a higher false refusal rate.\nLong-context safety. Long-context models are vulnerable to many-shot jailbreaking attacks without targeted\nmitigation (Anil et al., 2024). To address this, we finetune our models on SFT datasets that include examples\nof safe behavior in the presence of demonstrations of unsafe behavior in context. We develop a scalable\nmitigation strategy that significantly reduces VR, effectively neutralizing the impact of longer context attacks\neven for 256-shot attacks. This approach shows little to no impact on FRR and most helpfulness metrics.\nTo quantify the effectiveness of our long context safety mitigations, we use two additional benchmarking\nmethods: DocQA and Many-shot. For DocQA, short for “document question answering,” we use long documents\nwith information that could be utilized in adversarial ways. Models are provided both the document and a set\nof prompts related to the document in order to test whether the questions being related to information in the\ndocument affected the model’s ability to respond safely to the prompts. For Many-shot, following Anil et al.\n(2024), we construct a synthetic chat history composed of unsafe prompt-response pairs. A final prompt,\nunrelated to previous messages, is used to test whether the unsafe behavior in-context influenced the model\n45\nto response unsafely. The violation and false refusal rates for both DocQA and Many-shot are shown in\nFigure 20. We see that Llama 405B (with and without Llama Guard) is Pareto-better than the Comp. 2\nsystem across both violation rates and false refusal rates, across both DocQA and Many-shot. Relative to\nComp. 1, we find that Llama 405B is significantly safer, while coming at a trade off on false refusal.\nTool usage safety. The diversity of possible tools and the implementation of the tool usage call and integration\ninto the model make tool usage a challenging capability to fully mitigate (Wallace et al., 2024). We focus on\nthe search usecase. Violation and false refusal rates are shown in Figure 20. We tested against the Comp. 1\nsystem, where we find that Llama 405B is significantly safer, though has a slightly higher false refusal rate.\n5.4.5 Cybersecurity and Chemical/Biological Weapons Safety\nCyberSecurity evaluation results. To evaluate cybersecurity risk, we leverage the Cyber', document_id='num-0', token_count=512)0.7354530813978312
Chunk(content='.\nThrough careful ablations, we observe that mixing0.1% of synthetically generated long-context data with the\noriginal short-context data optimizes the performance across both short-context and long-context benchmarks.\nDPO. We observe that using only short context training data in DPO did not negatively impact long-context\nperformance as long as the SFT model is high quality in long context tasks. We suspect this is due to the\nfact that our DPO recipe has fewer optimizer steps than SFT. Given this finding, we keep the standard\nshort-context recipe for DPO on top of our long-context SFT checkpoints.\n4.3.5 Tool Use\nTeaching LLMs to use tools such as search engines or code interpreters hugely expands the range of tasks\nthey can solve, transforming them from pure chat models into more general assistants (Nakano et al., 2021;\nThoppilan et al., 2022; Parisi et al., 2022; Gao et al., 2023; Mialon et al., 2023a; Schick et al., 2024). We train\nLlama 3 to interact with the following tools:\n• Search engine. Llama 3 is trained to use Brave Search7 to answer questions about recent events that go\nbeyond its knowledge cutoff or that require retrieving a particular piece of information from the web.\n• Python interpreter. Llama 3 can generate and execute code to perform complex computations, read files\nuploaded by the user and solve tasks based on them such as question answering, summarization, data\nanalysis or visualization.\n7https://brave.com/search/api/\n24\n• Mathematical computational engine. Llama 3 can use the Wolfram Alpha API8 to more accurately solve\nmath, science problems, or retrieve accurate information from Wolfram’s database.\nThe resulting model is able to use these tools in a chat setup to solve the user’s queries, including in multi-turn\ndialogs. If a query requires multiple tool calls, the model can write a step-by-step plan, call the tools in\nsequence, and do reasoning after each tool call.\nWe also improve Llama 3’s zero-shot tool use capabilities — given in-context, potentially unseen tool definitions\nand a user query, we train the model to generate the correct tool call.\nImplementation. We implement our core tools as Python objects with different methods. Zero-shot tools can\nbe implemented as Python functions with descriptions, documentation (i.e., examples for', document_id='num-0', token_count=512)0.7350672465928054
Chunk(content=' Embeddings RoPE (θ = 500, 000)\nTable 3 Overview of the key hyperparameters of Llama 3. We display settings for 8B, 70B, and 405B language models.\n• We use a vocabulary with 128K tokens. Our token vocabulary combines 100K tokens from thetiktoken3\ntokenizer with 28K additional tokens to better support non-English languages. Compared to the Llama\n2 tokenizer, our new tokenizer improves compression rates on a sample of English data from 3.17 to\n3.94 characters per token. This enables the model to “read” more text for the same amount of training\ncompute. We also found that adding 28K tokens from select non-English languages improved both\ncompression ratios and downstream performance, with no impact on English tokenization.\n• We increase the RoPE base frequency hyperparameter to 500,000. This enables us to better support\nlonger contexts; Xiong et al. (2023) showed this value to be effective for context lengths up to 32,768.\nLlama 3 405B uses an architecture with 126 layers, a token representation dimension of 16,384, and 128\nattention heads; see Table 3 for details. This leads to a model size that is approximately compute-optimal\naccording to scaling laws on our data for our training budget of3.8 × 1025 FLOPs.\n3.2.1 Scaling Laws\nWe develop scaling laws (Hoffmann et al., 2022; Kaplan et al., 2020) to determine the optimal model size for\nour flagship model given our pre-training compute budget. In addition to determining the optimal model size,\na major challenge is to forecast the flagship model’s performance on downstream benchmark tasks, due to a\ncouple of issues: (1) Existing scaling laws typically predict only next-token prediction loss rather than specific\nbenchmark performance. (2) Scaling laws can be noisy and unreliable because they are developed based on\npre-training runs conducted with small compute budgets (Wei et al., 2022b).\nTo address these challenges, we implement a two-stage methodology to develop scaling laws that accurately\npredict downstream benchmark performance:\n1. We first establish a correlation between the compute-optimal model’s negative log-likelihood on down-\nstream tasks and the training FLOPs.\n2. Next, we correlate the negative log-likelihood on downstream tasks with task accuracy, utilizing both', document_id='num-0', token_count=512)0.7172908346230037
```
## Before submitting
- [x] N/A - This PR fixes a typo or improves the docs (you can dismiss
the other checks if that's the case).
- [x] Ran pre-commit to handle lint / formatting issues.
- [x] Read the [contributor
guideline](https://github.com/meta-llama/llama-stack/blob/main/CONTRIBUTING.md),
Pull Request section?
- [x] N/A - Updated relevant documentation.
- [x] Wrote necessary unit or integration tests.
# What does this PR do?
This PR moves all print statements to use logging. Things changed:
- Had to add `await start_trace("sse_generator")` to server.py to
actually get tracing working. else was not seeing any logs
- If no telemetry provider is provided in the run.yaml, we will write to
stdout
- by default, the logs are going to be in JSON, but we expose an option
to configure to output in a human readable way.
The semantics of an Update on resources is very tricky to reason about
especially for memory banks and models. The best way to go forward here
is for the user to unregister and register a new resource. We don't have
a compelling reason to support update APIs.
Tests:
pytest -v -s llama_stack/providers/tests/memory/test_memory.py -m
"chroma" --env CHROMA_HOST=localhost --env CHROMA_PORT=8000
pytest -v -s llama_stack/providers/tests/memory/test_memory.py -m
"pgvector" --env PGVECTOR_DB=postgres --env PGVECTOR_USER=postgres --env
PGVECTOR_PASSWORD=mysecretpassword --env PGVECTOR_HOST=0.0.0.0
$CONDA_PREFIX/bin/pytest -v -s -m "ollama"
llama_stack/providers/tests/inference/test_model_registration.py
---------
Co-authored-by: Dinesh Yeduguru <dineshyv@fb.com>
This PR adds support for Qdrant - https://qdrant.tech/ to be used as a vector memory.
I've unit-tested the methods to confirm that they work as intended.
To run Qdrant
```
docker run -p 6333:6333 qdrant/qdrant
```
This PR makes several core changes to the developer experience surrounding Llama Stack.
Background: PR #92 introduced the notion of "routing" to the Llama Stack. It introduces three object types: (1) models, (2) shields and (3) memory banks. Each of these objects can be associated with a distinct provider. So you can get model A to be inferenced locally while model B, C can be inference remotely (e.g.)
However, this had a few drawbacks:
you could not address the provider instances -- i.e., if you configured "meta-reference" with a given model, you could not assign an identifier to this instance which you could re-use later.
the above meant that you could not register a "routing_key" (e.g. model) dynamically and say "please use this existing provider I have already configured" for a new model.
the terms "routing_table" and "routing_key" were exposed directly to the user. in my view, this is way too much overhead for a new user (which almost everyone is.) people come to the stack wanting to do ML and encounter a completely unexpected term.
What this PR does: This PR structures the run config with only a single prominent key:
- providers
Providers are instances of configured provider types. Here's an example which shows two instances of the remote::tgi provider which are serving two different models.
providers:
inference:
- provider_id: foo
provider_type: remote::tgi
config: { ... }
- provider_id: bar
provider_type: remote::tgi
config: { ... }
Secondly, the PR adds dynamic registration of { models | shields | memory_banks } to the API surface. The distribution still acts like a "routing table" (as previously) except that it asks the backing providers for a listing of these objects. For example it asks a TGI or Ollama inference adapter what models it is serving. Only the models that are being actually served can be requested by the user for inference. Otherwise, the Stack server will throw an error.
When dynamically registering these objects, you can use the provider IDs shown above. Info about providers can be obtained using the Api.inspect set of endpoints (/providers, /routes, etc.)
The above examples shows the correspondence between inference providers and models registry items. Things work similarly for the safety <=> shields and memory <=> memory_banks pairs.
Registry: This PR also makes it so that Providers need to implement additional methods for registering and listing objects. For example, each Inference provider is now expected to implement the ModelsProtocolPrivate protocol (naming is not great!) which consists of two methods
register_model
list_models
The goal is to inform the provider that a certain model needs to be supported so the provider can make any relevant backend changes if needed (or throw an error if the model cannot be supported.)
There are many other cleanups included some of which are detailed in a follow-up comment.
This is yet another of those large PRs (hopefully we will have less and less of them as things mature fast). This one introduces substantial improvements and some simplifications to the stack.
Most important bits:
* Agents reference implementation now has support for session / turn persistence. The default implementation uses sqlite but there's also support for using Redis.
* We have re-architected the structure of the Stack APIs to allow for more flexible routing. The motivating use cases are:
- routing model A to ollama and model B to a remote provider like Together
- routing shield A to local impl while shield B to a remote provider like Bedrock
- routing a vector memory bank to Weaviate while routing a keyvalue memory bank to Redis
* Support for provider specific parameters to be passed from the clients. A client can pass data using `x_llamastack_provider_data` parameter which can be type-checked and provided to the Adapter implementations.
* API Keys passed from Client instead of distro configuration
* delete distribution registry
* Rename the "package" word away
* Introduce a "Router" layer for providers
Some providers need to be factorized and considered as thin routing
layers on top of other providers. Consider two examples:
- The inference API should be a routing layer over inference providers,
routed using the "model" key
- The memory banks API is another instance where various memory bank
types will be provided by independent providers (e.g., a vector store
is served by Chroma while a keyvalue memory can be served by Redis or
PGVector)
This commit introduces a generalized routing layer for this purpose.
* update `apis_to_serve`
* llama_toolchain -> llama_stack
* Codemod from llama_toolchain -> llama_stack
- added providers/registry
- cleaned up api/ subdirectories and moved impls away
- restructured api/api.py
- from llama_stack.apis.<api> import foo should work now
- update imports to do llama_stack.apis.<api>
- update many other imports
- added __init__, fixed some registry imports
- updated registry imports
- create_agentic_system -> create_agent
- AgenticSystem -> Agent
* Moved some stuff out of common/; re-generated OpenAPI spec
* llama-toolchain -> llama-stack (hyphens)
* add control plane API
* add redis adapter + sqlite provider
* move core -> distribution
* Some more toolchain -> stack changes
* small naming shenanigans
* Removing custom tool and agent utilities and moving them client side
* Move control plane to distribution server for now
* Remove control plane from API list
* no codeshield dependency randomly plzzzzz
* Add "fire" as a dependency
* add back event loggers
* stack configure fixes
* use brave instead of bing in the example client
* add init file so it gets packaged
* add init files so it gets packaged
* Update MANIFEST
* bug fix
---------
Co-authored-by: Hardik Shah <hjshah@fb.com>
Co-authored-by: Xi Yan <xiyan@meta.com>
Co-authored-by: Ashwin Bharambe <ashwin@meta.com>
