# What does this PR do?
the directory structure was src/llama-stack-api/llama_stack_api
instead it should just be src/llama_stack_api to match the other
packages.
update the structure and pyproject/linting config
---------
Signed-off-by: Charlie Doern <cdoern@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
Extract API definitions and provider specifications into a standalone
llama-stack-api package that can be published to PyPI independently of
the main llama-stack server.
see: https://github.com/llamastack/llama-stack/pull/2978 and
https://github.com/llamastack/llama-stack/pull/2978#issuecomment-3145115942
Motivation
External providers currently import from llama-stack, which overrides
the installed version and causes dependency conflicts. This separation
allows external providers to:
- Install only the type definitions they need without server
dependencies
- Avoid version conflicts with the installed llama-stack package
- Be versioned and released independently
This enables us to re-enable external provider module tests that were
previously blocked by these import conflicts.
Changes
- Created llama-stack-api package with minimal dependencies (pydantic,
jsonschema)
- Moved APIs, providers datatypes, strong_typing, and schema_utils
- Updated all imports from llama_stack.* to llama_stack_api.*
- Configured local editable install for development workflow
- Updated linting and type-checking configuration for both packages
Next Steps
- Publish llama-stack-api to PyPI
- Update external provider dependencies
- Re-enable external provider module tests
Pre-cursor PRs to this one:
- #4093
- #3954
- #4064
These PRs moved key pieces _out_ of the Api pkg, limiting the scope of
change here.
relates to #3237
## Test Plan
Package builds successfully and can be imported independently. All
pre-commit hooks pass with expected exclusions maintained.
---------
Signed-off-by: Charlie Doern <cdoern@redhat.com>
We would like to run all OpenAI compatibility tests using only the
openai-client library. This is most friendly for contributors since they
can run tests without needing to update the client-sdks (which is
getting easier but still a long pole.)
This is the first step in enabling that -- no using "library client" for
any of the Responses tests. This seems like a reasonable trade-off since
the usage of an embeddeble library client for Responses (or any
OpenAI-compatible) behavior seems to be not very common. To do this, we
needed to enable MCP tests (which only worked in library client mode)
for server mode.
Fixes issues in the storage system by guaranteeing immediate durability
for responses and ensuring background writers stay alive. Three related
fixes:
* Responses to the OpenAI-compatible API now write directly to
Postgres/SQLite inside the request instead of detouring through an async
queue that might never drain; this restores the expected
read-after-write behavior and removes the "response not found" races
reported by users.
* The access-control shim was stamping owner_principal/access_attributes
as SQL NULL, which Postgres interprets as non-public rows; fixing it to
use the empty-string/JSON-null pattern means conversations and responses
stored without an authenticated user stay queryable (matching SQLite).
* The inference-store queue remains for batching, but its worker tasks
now start lazily on the live event loop so server startup doesn't cancel
them—writes keep flowing even when the stack is launched via llama stack
run.
Closes#4115
### Test Plan
Added a matrix entry to test our "base" suite against Postgres as the
store.
Help users find the comprehensive integration testing docs by linking to
the record-replay documentation. This clarifies that the technical
README complements the main docs.
# What does this PR do?
- Updates `/vector_stores/{vector_store_id}/files/{file_id}/content` to
allow returning `embeddings` and `metadata` using the `extra_query`
- Updates the UI accordingly to display them.
- Update UI to support CRUD operations in the Vector Stores section and
adds a new modal exposing the functionality.
- Updates Vector Store update to fail if a user tries to update Provider
ID (which doesn't make sense to allow)
```python
In [1]: client.vector_stores.files.content(
vector_store_id=vector_store.id,
file_id=file.id,
extra_query={"include_embeddings": True, "include_metadata": True}
)
Out [1]: FileContentResponse(attributes={}, content=[Content(text='This is a test document to check if embeddings are generated properly.\n', type='text', embedding=[0.33760684728622437, ...,], chunk_metadata={'chunk_id': '62a63ae0-c202-f060-1b86-0a688995b8d3', 'document_id': 'file-27291dbc679642ac94ffac6d2810c339', 'source': None, 'created_timestamp': 1762053437, 'updated_timestamp': 1762053437, 'chunk_window': '0-13', 'chunk_tokenizer': 'DEFAULT_TIKTOKEN_TOKENIZER', 'chunk_embedding_model': 'sentence-transformers/nomic
-ai/nomic-embed-text-v1.5', 'chunk_embedding_dimension': 768, 'content_token_count': 13, 'metadata_token_count': 9}, metadata={'filename': 'test-embedding.txt', 'chunk_id': '62a63ae0-c202-f060-1b86-0a688995b8d3', 'document_id': 'file-27291dbc679642ac94ffac6d2810c339', 'token_count': 13, 'metadata_token_count': 9})], file_id='file-27291dbc679642ac94ffac6d2810c339', filename='test-embedding.txt')
```
Screenshots of UI are displayed below:
### List Vector Store with Added "Create New Vector Store"
<img width="1912" height="491" alt="Screenshot 2025-11-06 at 10 47
25 PM"
src="https://github.com/user-attachments/assets/a3a3ddd9-758d-4005-ac9c-5047f03916f3"
/>
### Create New Vector Store
<img width="1918" height="1048" alt="Screenshot 2025-11-06 at 10 47
49 PM"
src="https://github.com/user-attachments/assets/b4dc0d31-696f-4e68-b109-27915090f158"
/>
### Edit Vector Store
<img width="1916" height="1355" alt="Screenshot 2025-11-06 at 10 48
32 PM"
src="https://github.com/user-attachments/assets/ec879c63-4cf7-489f-bb1e-57ccc7931414"
/>
### Vector Store Files Contents page (with Embeddings)
<img width="1914" height="849" alt="Screenshot 2025-11-06 at 11 54
32 PM"
src="https://github.com/user-attachments/assets/3095520d-0e90-41f7-83bd-652f6c3fbf27"
/>
### Vector Store Files Contents Details page (with Embeddings)
<img width="1916" height="1221" alt="Screenshot 2025-11-06 at 11 55
00 PM"
src="https://github.com/user-attachments/assets/e71dbdc5-5b49-472b-a43a-5785f58d196c"
/>
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
## Test Plan
Tests added for Middleware extension and Provider failures.
---------
Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
# What does this PR do?
Add explicit connection cleanup and shorter timeouts to OpenAI client
fixtures. Fixes CI deadlock after 25+ tests due to connection pool
exhaustion. Also adds 60s timeout to test_conversation_context_loading
as safety net.
## Test Plan
tests pass
Signed-off-by: Charlie Doern <cdoern@redhat.com>
# Problem
Responses API uses max_tool_calls parameter to limit the number of tool
calls that can be generated in a response. Currently, LLS implementation
of the Responses API does not support this parameter.
# What does this PR do?
This pull request adds the max_tool_calls field to the response object
definition and updates the inline provider. it also ensures that:
- the total number of calls to built-in and mcp tools do not exceed
max_tool_calls
- an error is thrown if max_tool_calls < 1 (behavior seen with the
OpenAI Responses API, but we can change this if needed)
Closes #[3563](https://github.com/llamastack/llama-stack/issues/3563)
## Test Plan
- Tested manually for change in model response w.r.t supplied
max_tool_calls field.
- Added integration tests to test invalid max_tool_calls parameter.
- Added integration tests to check max_tool_calls parameter with
built-in and function tools.
- Added integration tests to check max_tool_calls parameter in the
returned response object.
- Recorded OpenAI Responses API behavior using a sample script:
https://github.com/s-akhtar-baig/llama-stack-examples/blob/main/responses/src/max_tool_calls.py
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
Adds OCI GenAI PaaS models for openai chat completion endpoints.
## Test Plan
In an OCI tenancy with access to GenAI PaaS, perform the following
steps:
1. Ensure you have IAM policies in place to use service (check docs
included in this PR)
2. For local development, [setup OCI
cli](https://docs.oracle.com/en-us/iaas/Content/API/SDKDocs/cliinstall.htm)
and configure the CLI with your region, tenancy, and auth
[here](https://docs.oracle.com/en-us/iaas/Content/API/SDKDocs/cliconfigure.htm)
3. Once configured, go through llama-stack setup and run llama-stack
(uses config based auth) like:
```bash
OCI_AUTH_TYPE=config_file \
OCI_CLI_PROFILE=CHICAGO \
OCI_REGION=us-chicago-1 \
OCI_COMPARTMENT_OCID=ocid1.compartment.oc1..aaaaaaaa5...5a \
llama stack run oci
```
4. Hit the `models` endpoint to list models after server is running:
```bash
curl http://localhost:8321/v1/models | jq
...
{
"identifier": "meta.llama-4-scout-17b-16e-instruct",
"provider_resource_id": "ocid1.generativeaimodel.oc1.us-chicago-1.am...q",
"provider_id": "oci",
"type": "model",
"metadata": {
"display_name": "meta.llama-4-scout-17b-16e-instruct",
"capabilities": [
"CHAT"
],
"oci_model_id": "ocid1.generativeaimodel.oc1.us-chicago-1.a...q"
},
"model_type": "llm"
},
...
```
5. Use the "display_name" field to use the model in a
`/chat/completions` request:
```bash
# Streaming result
curl -X POST http://localhost:8321/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "meta.llama-4-scout-17b-16e-instruct",
"stream": true,
"temperature": 0.9,
"messages": [
{
"role": "system",
"content": "You are a funny comedian. You can be crass."
},
{
"role": "user",
"content": "Tell me a funny joke about programming."
}
]
}'
# Non-streaming result
curl -X POST http://localhost:8321/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "meta.llama-4-scout-17b-16e-instruct",
"stream": false,
"temperature": 0.9,
"messages": [
{
"role": "system",
"content": "You are a funny comedian. You can be crass."
},
{
"role": "user",
"content": "Tell me a funny joke about programming."
}
]
}'
```
6. Try out other models from the `/models` endpoint.
# What does this PR do?
Resolves#4102
1. Added `web_search_2025_08_26` to the `WebSearchToolTypes` list and
the `OpenAIResponseInputToolWebSearch.type` Literal union
2. No changes needed to tool execution logic - all `web_search` types
map to the same underlying tool
3. Backward compatibility is maintained - existing `web_search`,
`web_search_preview`, and `web_search_preview_2025_03_11` types continue
to work
4. Added an integration test case using {"type":
"web_search_2025_08_26"} to verify it works correctly
5. Updated `docs/docs/providers/openai_responses_limitations.mdx` to
reflect that `web_search_2025_08_26` is now supported.
6. Removed incorrect references to `MOD1/MOD2/MOD3` (which don't exist
in the codebase)
<!-- 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.* -->
---------
Signed-off-by: Aakanksha Duggal <aduggal@redhat.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
This dependency has been bothering folks for a long time (cc @leseb). We
really needed it due to "library client" which is primarily used for our
tests and is not a part of the Stack server. Anyone who needs to use the
library client can certainly install `llama-stack-client` in their
environment to make that work.
Updated the notebook references to install `llama-stack-client`
additionally when setting things up.
https://github.com/llamastack/llama-stack/pull/4055 cleaned the agents
implementation but while doing so it removed some tests which actually
corresponded to the responses implementation. This PR brings those tests
and assocated recordings back.
(We should likely combine all responses tests into one suite, but that
is beyond the scope of this PR.)
o Introduces vLLM provider support to the record/replay testing
framework
o Enabling both recording and replay of vLLM API interactions alongside
existing Ollama support.
The changes enable testing of vLLM functionality. vLLM tests focus on
inference capabilities, while Ollama continues to exercise the full API
surface
including vision features.
--
This is an alternative to #3128 , using qwen3 instead of llama 3.2 1B
appears to be more capable at structure output and tool calls.
---------
Signed-off-by: Derek Higgins <derekh@redhat.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
# What does this PR do?
- when create vector store is called without chunk strategy, we actually
the strategy used so that the value is persisted instead of
strategy='None'
## Test Plan
updated tests
# What does this PR do?
1. Make telemetry tests as easy as possible for users by expanding the
`SpanStub` data class and creating the `MetricStub` dataclass as a way
to consistently marshal telemetry data in test fixtures and unmarshal
and handle it in tests.
2. Structure server and client tests to always follow the same standards
for consistent testing experience by using the `SpanStub` and
`MetricStub` data class objects.
3. Enable Metrics Testing for completions endpoint
4. Correct token metrics to use histograms instead of counts to capture
tokens per request rather than a cumulative count of tokens over the
lifecycle of the server.
## Test Plan
These are tests
Added a script to cleanup recordings. While doing this, moved the CI
matrix generation to a separate script so there is a single source of
truth for the matrix.
Ran the cleanup script as:
```
PYTHONPATH=. python scripts/cleanup_recordings.py
```
Also added this as part of the pre-commit workflow to ensure that the
recordings are always up to date and that no stale recordings are left
in the repo.
- Removes the deprecated agents (sessions and turns) API that was marked
alpha in 0.3.0
- Cleans up unused imports and orphaned types after the API removal
- Removes `SessionNotFoundError` and `AgentTurnInputType` which are no
longer needed
The agents API is completely superseded by the Responses + Conversations
APIs, and the client SDK Agent class already uses those implementations.
Corresponding client-side PR:
https://github.com/llamastack/llama-stack-client-python/pull/295
The llama-stack-client now uses /`v1/openai/v1/models` which returns
OpenAI-compatible model objects with 'id' and 'custom_metadata' fields
instead of the Resource-style 'identifier' field. Updated api_recorder
to handle the new endpoint and modified tests to access model metadata
appropriately. Deleted stale model recordings for re-recording.
**NOTE: CI will be red on this one since it is dependent on
https://github.com/llamastack/llama-stack-client-python/pull/291/files
landing. I verified locally that it is green.**
Without this hint Qwen3-0.6B tends to reply with the full name
and sometimes doesn't reply with the correct drafted year.
---------
Signed-off-by: Derek Higgins <derekh@redhat.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
Allow filtering for v1alpha, v1beta, deprecated and v1. Backward
incompatible change since by default it only returns v1 apis now.
## Test Plan
added unit test
# What does this PR do?
Add rerank API for NVIDIA Inference Provider.
<!-- If resolving an issue, uncomment and update the line below -->
Closes#3278
## Test Plan
Unit test:
```
pytest tests/unit/providers/nvidia/test_rerank_inference.py
```
Integration test:
```
pytest -s -v tests/integration/inference/test_rerank.py --stack-config="inference=nvidia" --rerank-model=nvidia/nvidia/nv-rerankqa-mistral-4b-v3 --env NVIDIA_API_KEY="" --env NVIDIA_BASE_URL="https://integrate.api.nvidia.com"
```
… case variations
The ollama/llama3.2:3b-instruct-fp16 model returns string values with
trailing whitespace in structured JSON output. Updated test assertions
to use case-insensitive substring matching instead of exact equality.
Use .lower() for case-insensitive comparison
Check if expected value is contained in actual value (handles
whitespace)
Closes: #3996
Signed-off-by: Derek Higgins <derekh@redhat.com>
This should be "remote::vllm". This causes some log probs tests to be
skipped with remote vllm. (They
fail if run).
Signed-off-by: Derek Higgins <derekh@redhat.com>
# What does this PR do?
chunk_id in the Chunk class executes actual logic to compute a chunk ID.
This sort of logic should not live in the API spec.
Instead, the providers should be in charge of calling generate_chunk_id,
and pass it to `Chunk`.
this removes the incorrect dependency between Provider impl and API impl
Signed-off-by: Charlie Doern <cdoern@redhat.com>
# What does this PR do?
- Adds OpenAI files provider
- Note that file content retrieval is pretty limited by `purpose`
https://community.openai.com/t/file-uploads-error-why-can-t-i-download-files-with-purpose-user-data/1357013?utm_source=chatgpt.com
## Test Plan
Modify run yaml to use openai files provider:
```
files:
- provider_id: openai
provider_type: remote::openai
config:
api_key: ${env.OPENAI_API_KEY:=}
metadata_store:
backend: sql_default
table_name: openai_files_metadata
# Then run files tests
❯ uv run --no-sync ./scripts/integration-tests.sh --stack-config server:ci-tests --inference-mode replay --setup ollama --suite base --pattern test_files
```
This PR enables routing of fully qualified model IDs of the form
`provider_id/model_id` even when the models are not registered with the
Stack.
Here's the situation: assume a remote inference provider which works
only when users provide their own API keys via
`X-LlamaStack-Provider-Data` header. By definition, we cannot list
models and hence update our routing registry. But because we _require_ a
provider ID in the models now, we can identify which provider to route
to and let that provider decide.
Note that we still try to look up our registry since it may have a
pre-registered alias. Just that we don't outright fail when we are not
able to look it up.
Also, updated inference router so that the responses have the _exact_
model that the request had.
## Test Plan
Added an integration test
Closes#3929
---------
Co-authored-by: ehhuang <ehhuang@users.noreply.github.com>
This patch ensures if max tokens is not defined, then is set to None
instead of 0 when calling openai_chat_completion. This way some
providers (like gemini) that cannot handle the `max_tokens = 0` will not
fail
Issue: #3666
The vector_provider_wrapper was only limiting providers to
faiss/sqlite-vec for replay mode, but CI tests also run in record mode
with the same limited set of providers. This caused test failures when
trying to test against milvus, chromadb, pgvector, weaviate, and qdrant
which aren't configured in the record job.
# What does this PR do?
Clean up telemetry code since the telemetry API has been remove.
- moved telemetry files out of providers to core
- removed from Api
## Test Plan
❯ OTEL_SERVICE_NAME=llama_stack
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 uv run llama stack run
starter
❯ curl http://localhost:8321/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-4o-mini",
"messages": [
{
"role": "user",
"content": "Hello!"
}
]
}'
-> verify traces in Grafana
CI
Let us enable responses suite in CI now.
Also a minor fix: MCP tool tests intentionally trigger authentication
failures to verify error handling, but the resulting error logs clutter
test output.
Move conversation sync logic before yield to ensure it executes even
when
streaming consumers break early after receiving response.completed
event.
## Test Plan
```
OLLAMA_URL=http://localhost:11434 \
pytest -sv tests/integration/responses/ \
--stack-config server:ci-tests \
--text-model ollama/llama3.2:3b-instruct-fp16 \
--inference-mode live \
-k conversation_multi
```
This test now passes.
# What does this PR do?
Refactor setting default vector store provider and embedding model to
use an optional `vector_stores` config in the `StackRunConfig` and clean
up code to do so (had to add back in some pieces of VectorDB). Also
added remote Qdrant and Weaviate to starter distro (based on other PR
where inference providers were added for UX).
New config is simply (default for Starter distro):
```yaml
vector_stores:
default_provider_id: faiss
default_embedding_model:
provider_id: sentence-transformers
model_id: nomic-ai/nomic-embed-text-v1.5
```
## Test Plan
CI and Unit tests.
---------
Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
**This PR changes configurations in a backward incompatible way.**
Run configs today repeat full SQLite/Postgres snippets everywhere a
store is needed, which means duplicated credentials, extra connection
pools, and lots of drift between files. This PR introduces named storage
backends so the stack and providers can share a single catalog and
reference those backends by name.
## Key Changes
- Add `storage.backends` to `StackRunConfig`, register each KV/SQL
backend once at startup, and validate that references point to the right
family.
- Move server stores under `storage.stores` with lightweight references
(backend + namespace/table) instead of full configs.
- Update every provider/config/doc to use the new reference style;
docs/codegen now surface the simplified YAML.
## Migration
Before:
```yaml
metadata_store:
type: sqlite
db_path: ~/.llama/distributions/foo/registry.db
inference_store:
type: postgres
host: ${env.POSTGRES_HOST}
port: ${env.POSTGRES_PORT}
db: ${env.POSTGRES_DB}
user: ${env.POSTGRES_USER}
password: ${env.POSTGRES_PASSWORD}
conversations_store:
type: postgres
host: ${env.POSTGRES_HOST}
port: ${env.POSTGRES_PORT}
db: ${env.POSTGRES_DB}
user: ${env.POSTGRES_USER}
password: ${env.POSTGRES_PASSWORD}
```
After:
```yaml
storage:
backends:
kv_default:
type: kv_sqlite
db_path: ~/.llama/distributions/foo/kvstore.db
sql_default:
type: sql_postgres
host: ${env.POSTGRES_HOST}
port: ${env.POSTGRES_PORT}
db: ${env.POSTGRES_DB}
user: ${env.POSTGRES_USER}
password: ${env.POSTGRES_PASSWORD}
stores:
metadata:
backend: kv_default
namespace: registry
inference:
backend: sql_default
table_name: inference_store
max_write_queue_size: 10000
num_writers: 4
conversations:
backend: sql_default
table_name: openai_conversations
```
Provider configs follow the same pattern—for example, a Chroma vector
adapter switches from:
```yaml
providers:
vector_io:
- provider_id: chromadb
provider_type: remote::chromadb
config:
url: ${env.CHROMADB_URL}
kvstore:
type: sqlite
db_path: ~/.llama/distributions/foo/chroma.db
```
to:
```yaml
providers:
vector_io:
- provider_id: chromadb
provider_type: remote::chromadb
config:
url: ${env.CHROMADB_URL}
persistence:
backend: kv_default
namespace: vector_io::chroma_remote
```
Once the backends are declared, everything else just points at them, so
rotating credentials or swapping to Postgres happens in one place and
the stack reuses a single connection pool.
# Problem
The current inline provider appends the user provided instructions to
messages as a system prompt, but the returned response object does not
contain the instructions field (as specified in the OpenAI responses
spec).
# What does this PR do?
This pull request adds the instruction field to the response object
definition and updates the inline provider. It also ensures that
instructions from previous response is not carried over to the next
response (as specified in the openAI spec).
Closes #[3566](https://github.com/llamastack/llama-stack/issues/3566)
## Test Plan
- Tested manually for change in model response w.r.t supplied
instructions field.
- Added unit test to check that the instructions from previous response
is not carried over to the next response.
- Added integration tests to check instructions parameter in the
returned response object.
- Added new recordings for the integration tests.
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
# What does this PR do?
Adds a test and a standardized way to build future tests out for
telemetry in llama stack.
Contributes to https://github.com/llamastack/llama-stack/issues/3806
## Test Plan
This is the test plan 😎
In replay mode, inference is instantenous. We don't need to wait 15
seconds for the batch to be done. Fixing polling to do exp backoff makes
things work super fast.
