# Design Document: Remote vLLM Support

## Overview

This design introduces an LLM provider abstraction layer into Stonks Oracle so that both the existing Ollama backend and a new remote vLLM backend can be used interchangeably for document extraction and event classification. The vLLM server at `http://192.168.42.254:8000` runs `RedHatAI/Qwen3.6-35B-A3B-NVFP4` on an NVIDIA RTX 5090 with tensor parallelism and exposes an OpenAI-compatible `/v1/chat/completions` API.

The design preserves full backward compatibility — existing Ollama deployments work without any configuration changes. Provider selection is driven by the existing `model_provider` column in the `ai_agents` and `agent_variants` database tables, requiring no new migrations.

## Architecture

```mermaid
graph TD
    subgraph "Extractor Worker"
        MAIN[main.py]
        FACTORY[LLMClientFactory]
        EXTRACT[Extraction Pipeline]
        CLASSIFY[Event Classification Pipeline]
    end

    subgraph "Provider Abstraction"
        PROTO[LLMClient Protocol]
        OLLAMA_IMPL[OllamaClient]
        VLLM_IMPL[VLLMClient]
    end

    subgraph "Configuration"
        RESOLVER[AgentConfigResolver]
        OLLAMA_CFG[OllamaConfig]
        VLLM_CFG[VLLMConfig]
        APP_CFG[AppConfig]
    end

    subgraph "External Services"
        OLLAMA_SRV[Ollama Server<br/>:11434/api/chat]
        VLLM_SRV[vLLM Server<br/>:8000/v1/chat/completions]
    end

    MAIN --> FACTORY
    FACTORY --> PROTO
    PROTO --> OLLAMA_IMPL
    PROTO --> VLLM_IMPL
    EXTRACT --> PROTO
    CLASSIFY --> PROTO

    RESOLVER --> FACTORY
    OLLAMA_CFG --> FACTORY
    VLLM_CFG --> FACTORY
    APP_CFG --> OLLAMA_CFG
    APP_CFG --> VLLM_CFG

    OLLAMA_IMPL --> OLLAMA_SRV
    VLLM_IMPL --> VLLM_SRV
```

The key architectural decision is to use a Python `Protocol` (structural typing) rather than an ABC for the LLM client interface. This allows the existing `OllamaClient` to satisfy the protocol without inheritance changes, maintaining backward compatibility. The `VLLMClient` is a new class that also satisfies the protocol.

A factory function in `services/extractor/llm_factory.py` takes a `ResolvedAgentConfig` and the base configs, returning the appropriate client. The extractor worker (`main.py`) uses this factory instead of directly constructing `OllamaClient`.

## Components and Interfaces

### 1. LLM Client Protocol (`services/shared/llm_protocol.py`)

A `typing.Protocol` defining the contract both clients must satisfy:

```python
from typing import Protocol, runtime_checkable

@runtime_checkable
class LLMClient(Protocol):
    async def call_llm(
        self,
        prompts: dict[str, str],
        json_schema: dict[str, object],
        document_text: str = "",
    ) -> "ExtractionAttempt": ...

    async def close(self) -> None: ...
```

The `call_llm` method signature matches the existing `OllamaClient._call_ollama()` parameters and return type. The `OllamaClient` gains a public `call_llm` method that delegates to `_call_ollama()`, preserving the private method for internal backward compatibility.

### 2. VLLMClient (`services/extractor/vllm_client.py`)

New client implementing the `LLMClient` protocol for the OpenAI-compatible API:

```python
@dataclass
class VLLMClient:
    _config: VLLMConfig
    _http: httpx.AsyncClient
    _owns_client: bool

    async def call_llm(
        self,
        prompts: dict[str, str],
        json_schema: dict[str, object],
        document_text: str = "",
    ) -> ExtractionAttempt: ...

    async def close(self) -> None: ...
```

**Request format** (OpenAI-compatible):
```json
{
    "model": "RedHatAI/Qwen3.6-35B-A3B-NVFP4",
    "messages": [
        {"role": "system", "content": "..."},
        {"role": "user", "content": "..."}
    ],
    "max_tokens": 4096,
    "temperature": 0.7,
    "response_format": {"type": "json_object"}
}
```

**Response parsing**: Extracts `choices[0].message.content`, then applies the same `_strip_markdown_fences()` and `_repair_json()` pipeline as `OllamaClient`.

**Error handling**: Maps HTTP errors to the same string format as `OllamaClient` (`timeout`, `http_{code}`, `connection_error: {details}`, `empty_model_response`), so the existing `_is_retryable()` function works without modification.

**Key differences from OllamaClient**:
- Endpoint: `/v1/chat/completions` instead of `/api/chat`
- No `think: false`, `stream: false`, or `options` block
- Uses `max_tokens` instead of `options.num_predict`
- Uses `response_format: {"type": "json_object"}` for structured output
- Supports `temperature` parameter (Ollama uses model defaults)
- Response in `choices[0].message.content` instead of `message.content`

### 3. VLLMConfig (`services/shared/config.py`)

New dataclass alongside `OllamaConfig`:

```python
@dataclass
class VLLMConfig:
    base_url: str = "http://192.168.42.254:8000"
    model: str = "RedHatAI/Qwen3.6-35B-A3B-NVFP4"
    timeout: int = 120
    max_retries: int = 2
    retry_base_delay: float = 1.0
    retry_max_delay: float = 10.0
    retry_backoff_multiplier: float = 2.0
    max_tokens: int = 32768
    temperature: float = 0.7
    api_key: str = ""  # Optional, for authenticated vLLM deployments
```

Loaded from `VLLM_*` environment variables in `load_config()`. Added to `AppConfig` as `vllm: VLLMConfig`.

### 4. LLM Client Factory (`services/extractor/llm_factory.py`)

Factory function that replaces the hardcoded `OllamaClient` construction:

```python
def build_llm_client(
    resolved: ResolvedAgentConfig | None,
    ollama_config: OllamaConfig,
    vllm_config: VLLMConfig,
    http_client: httpx.AsyncClient | None = None,
) -> LLMClient:
    """Return the appropriate LLM client based on resolved provider."""
    ...

def build_config_from_resolved(
    resolved: ResolvedAgentConfig,
    base_ollama: OllamaConfig,
    base_vllm: VLLMConfig,
) -> OllamaConfig | VLLMConfig:
    """Build provider-specific config from resolved agent config."""
    ...
```

Provider routing logic:
1. If `resolved` is `None` or `resolved.model_provider` is `"ollama"` or empty → `OllamaClient`
2. If `resolved.model_provider` is `"vllm"` → `VLLMClient`
3. Unknown provider → log warning, fall back to `OllamaClient`

### 5. Updated Extractor Worker (`services/extractor/main.py`)

Changes to `main()`:
- Replace `_build_ollama_config_from_resolved()` with `build_llm_client()` from the factory
- Store clients as `LLMClient` type instead of `OllamaClient`
- On config refresh (every 100 jobs), detect provider changes and swap clients
- Log provider switches at INFO level

Changes to `_process_macro_classification()`:
- Accept `LLMClient` instead of `OllamaClient` for the classifier parameter

### 6. Updated OllamaClient (`services/extractor/client.py`)

Minimal changes to satisfy the protocol:
- Add public `call_llm()` method that delegates to `_call_ollama()`
- Keep `_call_ollama()` as-is for backward compatibility
- The `extract()` method continues to call `_call_ollama()` internally

### 7. Updated Event Classifier (`services/extractor/event_classifier.py`)

Changes to `classify_global_event()`:
- Accept `LLMClient` instead of `Any` for the `ollama_client` parameter
- Call `client.call_llm()` instead of `ollama_client._call_ollama()`
- Set `ModelMetadata.provider` based on the actual client type (inspect `_config` or pass provider string)

### 8. Helm Values (`infra/helm/stonks-oracle/values.yaml`)

New config entries:
```yaml
config:
  VLLM_BASE_URL: "http://192.168.42.254:8000"
  VLLM_MODEL: "RedHatAI/Qwen3.6-35B-A3B-NVFP4"
  VLLM_TIMEOUT: "120"
  VLLM_MAX_RETRIES: "2"
  VLLM_TEMPERATURE: "0.7"
  VLLM_API_KEY: ""
```

### 9. Health Check (`services/extractor/vllm_client.py`)

Startup validation function:

```python
async def check_vllm_health(base_url: str, timeout: float = 10.0) -> bool:
    """GET {base_url}/v1/models to verify vLLM is reachable."""
    ...
```

Called from `main()` when the resolved or default config specifies vLLM. On failure, logs WARNING and falls back to Ollama. On success, logs INFO with server URL and model list.

## Data Models

### VLLMConfig Dataclass

| Field | Type | Default | Env Var |
|-------|------|---------|---------|
| `base_url` | `str` | `http://192.168.42.254:8000` | `VLLM_BASE_URL` |
| `model` | `str` | `RedHatAI/Qwen3.6-35B-A3B-NVFP4` | `VLLM_MODEL` |
| `timeout` | `int` | `120` | `VLLM_TIMEOUT` |
| `max_retries` | `int` | `2` | `VLLM_MAX_RETRIES` |
| `retry_base_delay` | `float` | `1.0` | `VLLM_RETRY_BASE_DELAY` |
| `retry_max_delay` | `float` | `10.0` | `VLLM_RETRY_MAX_DELAY` |
| `retry_backoff_multiplier` | `float` | `2.0` | `VLLM_RETRY_BACKOFF_MULTIPLIER` |
| `max_tokens` | `int` | `32768` | `VLLM_MAX_TOKENS` |
| `temperature` | `float` | `0.7` | `VLLM_TEMPERATURE` |
| `api_key` | `str` | `""` | `VLLM_API_KEY` |

### ExtractionAttempt (unchanged)

The existing `ExtractionAttempt` dataclass is reused as-is for both providers. No changes needed.

### ModelMetadata (unchanged structure, new values)

The `provider` field now accepts `"vllm"` in addition to `"ollama"`. No schema change needed.

## Error Handling

### Error String Format Parity

Both clients produce identical error string formats so `_is_retryable()` works unchanged:

| Condition | Error String | Retryable |
|-----------|-------------|-----------|
| HTTP timeout | `timeout` | Yes |
| HTTP 400/401/403/404/422 | `http_{code}` | No |
| HTTP 500/502/503/429 | `http_{code}` | Yes |
| Connection refused/reset | `connection_error: {details}` | Yes |
| Empty response body | `empty_model_response` | Yes |
| Invalid JSON in response | `invalid_response_json` | Yes |

### Health Check Failure

If the vLLM health check fails at startup:
1. Log WARNING with the error details
2. Fall back to `OllamaClient` using `OllamaConfig`
3. Continue operation — the system degrades gracefully rather than crashing

### Provider Switch During Refresh

When the config refresh (every 100 jobs) detects a provider change:
1. Close the old client (`await old_client.close()`)
2. Construct the new client via the factory
3. Log the switch at INFO level
4. If new client construction fails, keep the old client and log ERROR

## Testing Strategy

### Property-Based Tests (`tests/test_pbt_llm_provider.py`)

Property-based tests using Hypothesis to verify the provider abstraction:

**P1: Provider factory routing property** (Req 3.4, 3.5, 9.5)
For all `model_provider` values in `{"ollama", "vllm", "", None}`, the factory returns the correct client type. For `"ollama"`, empty, or `None`, returns `OllamaClient`. For `"vllm"`, returns `VLLMClient`.

**P2: Error string format consistency property** (Req 5.6)
For all HTTP status codes (100-599), both `OllamaClient` and `VLLMClient` produce error strings in the same format (`http_{code}`), and `_is_retryable()` returns the same result for both.

**P3: VLLMClient request payload structure property** (Req 2.1, 8.1)
For all generated prompt dicts (system + user messages of arbitrary text), the VLLMClient produces a request payload that: contains `model`, `messages`, `max_tokens`, `temperature`; does NOT contain `think`, `stream`, `options`, `num_ctx`, `num_predict`.

**P4: JSON repair idempotence property** (Req 2.4)
For all valid JSON strings, `_repair_json(json_str)` returns a string that `json.loads()` can parse, and `_repair_json(_repair_json(json_str)) == _repair_json(json_str)` (idempotence).

**P5: Markdown fence stripping round-trip property** (Req 2.3)
For all strings `s`, `_strip_markdown_fences(f"```json\n{s}\n```")` returns `s` (stripped), and `_strip_markdown_fences(s)` returns `s` when no fences are present (identity).

**P6: VLLMConfig default construction property** (Req 3.1)
For all VLLMConfig instances constructed with default values, `base_url` is non-empty, `timeout > 0`, `max_retries >= 0`, `temperature` is between 0.0 and 2.0, and `max_tokens > 0`.

### Unit Tests (`tests/test_vllm_client.py`)

Example-based tests for specific behaviors:

- VLLMClient sends correct payload to `/v1/chat/completions` (mock httpx)
- VLLMClient extracts content from `choices[0].message.content`
- VLLMClient handles empty choices array → `empty_model_response`
- VLLMClient handles timeout → `timeout` error
- VLLMClient handles HTTP 500 → `http_500` error, retryable
- VLLMClient handles HTTP 400 → `http_400` error, non-retryable
- VLLMClient handles connection refused → `connection_error: ...`
- VLLMClient applies markdown fence stripping
- VLLMClient applies JSON repair
- VLLMClient includes temperature in payload
- VLLMClient includes `response_format` in payload
- Health check success logs INFO
- Health check failure logs WARNING and returns False
- Factory returns OllamaClient for provider="ollama"
- Factory returns VLLMClient for provider="vllm"
- Factory returns OllamaClient for provider="" (default)
- Factory returns OllamaClient for unknown provider with warning
- VLLMConfig loads from environment variables
- AppConfig includes vllm field with defaults
- OllamaClient.call_llm() delegates to _call_ollama()

### Existing Tests (unchanged)

- `tests/test_ollama_client.py` — continues to pass without modification
- All other existing test files — unaffected

## File Changes Summary

| File | Change Type | Description |
|------|-------------|-------------|
| `services/shared/llm_protocol.py` | **New** | `LLMClient` Protocol definition |
| `services/extractor/vllm_client.py` | **New** | `VLLMClient` implementation + health check |
| `services/extractor/llm_factory.py` | **New** | Factory function for provider routing |
| `services/shared/config.py` | **Modified** | Add `VLLMConfig`, update `AppConfig`, update `load_config()` |
| `services/extractor/client.py` | **Modified** | Add `call_llm()` public method to `OllamaClient` |
| `services/extractor/event_classifier.py` | **Modified** | Use `call_llm()` instead of `_call_ollama()`, accept `LLMClient` type |
| `services/extractor/main.py` | **Modified** | Use factory, support provider switching, health check |
| `infra/helm/stonks-oracle/values.yaml` | **Modified** | Add `VLLM_*` config entries |
| `tests/test_pbt_llm_provider.py` | **New** | Property-based tests for provider abstraction |
| `tests/test_vllm_client.py` | **New** | Unit tests for VLLMClient and factory |