Celes Renata
a72f336ad1
Multi-stage evidence-grounded inference architecture replacing the monolithic 9B model extraction pipeline. CPU-first specialist services handle routine extraction while the 9B vLLM model is preserved for semantic adjudication of ambiguous cases. Key components: - Capability-aware inference gateway (OpenAI-compatible + Ollama) - Endpoint registry with DB migrations and REST API - Sentence-aware document segmenter (property tests) - Deterministic financial parsing with offset integrity - Symbol resolution with ambiguity detection - Specialist service (GLiNER2, dynamic batching, K8s deployment) - Company-specific sentiment (FinBERT, calibration) - Retrieval-based novelty and duplicate detection - Confidence calibration pipeline - Deterministic routing engine (property tests) - 9B adjudication layer with VRAM gating - Stock-specific impact model (features, labels, baseline, trained) - Pipeline orchestrator (state machine, queues, leases, feature flags) - Bounded parallelism (async workers, semaphore, load shedding) - Observability (tracing, metrics, alerts) - Compatibility adapter (v3→v2 golden mapping tests) - Shadow/canary promotion framework - Active learning and fine-tuning pipeline Test results: 1,161 tests pass, ruff lint clean. All 282 spec tasks completed.
383 lines
14 KiB
Python
383 lines
14 KiB
Python
"""Ollama native client for the shared inference gateway.
|
|
|
|
Implements generation via the Ollama /api/chat endpoint with:
|
|
- Shared StructuredGenerationRequest/InferenceResult types
|
|
- Native JSON schema formatting when supported (format field)
|
|
- Prompt-only fallback with explicit reporting
|
|
- Stall/loop detection as Ollama-specific policy
|
|
- Configurable max output tokens (num_predict) and context window (num_ctx)
|
|
- Error mapping to shared InferenceErrorCategory
|
|
|
|
Requirements: 2.1, 2.6
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import logging
|
|
import time
|
|
from dataclasses import dataclass
|
|
|
|
import httpx
|
|
|
|
from services.shared.inference.errors import InferenceError, InferenceErrorCategory
|
|
from services.shared.inference.models import (
|
|
InferenceResult,
|
|
InferenceTarget,
|
|
StructuredGenerationRequest,
|
|
TokenUsage,
|
|
)
|
|
|
|
# Re-export for backward compatibility
|
|
__all__ = ["OllamaNativeClient", "StallPolicy"]
|
|
|
|
logger = logging.getLogger("inference.ollama_native")
|
|
|
|
|
|
@dataclass
|
|
class StallPolicy:
|
|
"""Ollama-specific stall/loop detection configuration.
|
|
|
|
Monitors streaming responses for repetitive output patterns.
|
|
This policy is intentionally Ollama-specific and does not leak
|
|
into the generic InferenceResult interface.
|
|
"""
|
|
|
|
enabled: bool = True
|
|
check_interval_seconds: float = 5.0
|
|
max_unchanged_intervals: int = 6
|
|
loop_window: int = 64
|
|
loop_threshold: float = 0.5
|
|
|
|
|
|
def _map_http_status_to_category(status_code: int) -> InferenceErrorCategory:
|
|
"""Map Ollama HTTP status codes to normalized error categories."""
|
|
if status_code == 401:
|
|
return InferenceErrorCategory.AUTH_FAILED
|
|
if status_code == 403:
|
|
return InferenceErrorCategory.FORBIDDEN
|
|
if status_code == 404:
|
|
return InferenceErrorCategory.MODEL_NOT_FOUND
|
|
if status_code == 429:
|
|
return InferenceErrorCategory.RATE_LIMITED
|
|
if status_code in (400, 422):
|
|
return InferenceErrorCategory.BAD_REQUEST
|
|
if status_code >= 500:
|
|
return InferenceErrorCategory.SERVER_ERROR
|
|
return InferenceErrorCategory.UNKNOWN
|
|
|
|
|
|
def _detect_loop(content: str, window: int, threshold: float) -> bool:
|
|
"""Detect repetitive output by checking if the tail repeats earlier content.
|
|
|
|
Looks at the last `window` characters and checks if a significant
|
|
portion of the tail matches an earlier substring, indicating the model
|
|
is stuck in a generation loop.
|
|
"""
|
|
if len(content) < window * 2:
|
|
return False
|
|
|
|
tail = content[-window:]
|
|
body = content[:-window]
|
|
|
|
# Check if the tail appears verbatim in the preceding content
|
|
if tail in body:
|
|
return True
|
|
|
|
# Check character-level repetition ratio in the tail
|
|
if not tail:
|
|
return False
|
|
unique_chars = len(set(tail))
|
|
ratio = unique_chars / len(tail)
|
|
return ratio < threshold
|
|
|
|
|
|
class OllamaNativeClient:
|
|
"""Async client for Ollama /api/chat using shared inference types.
|
|
|
|
Translates StructuredGenerationRequest into Ollama's native API format
|
|
and returns InferenceResult with proper metadata and error classification.
|
|
|
|
Stall/loop detection is an Ollama-specific policy that monitors streaming
|
|
responses for repetitive patterns and aborts generation when detected.
|
|
This does not affect the InferenceResult interface — stall detection
|
|
raises an InferenceError with category STALL_DETECTED.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
target: InferenceTarget,
|
|
*,
|
|
http_client: httpx.AsyncClient | None = None,
|
|
stall_policy: StallPolicy | None = None,
|
|
) -> None:
|
|
if target.protocol != "ollama_native":
|
|
raise ValueError(
|
|
f"OllamaNativeClient requires protocol='ollama_native', got '{target.protocol}'"
|
|
)
|
|
self._target = target
|
|
self._stall_policy = stall_policy or StallPolicy()
|
|
self._owns_client = http_client is None
|
|
self._http = http_client or httpx.AsyncClient(
|
|
timeout=httpx.Timeout(300.0, read=300.0),
|
|
)
|
|
|
|
async def close(self) -> None:
|
|
"""Close the underlying HTTP client if we own it."""
|
|
if self._owns_client:
|
|
await self._http.aclose()
|
|
|
|
async def generate(self, request: StructuredGenerationRequest) -> InferenceResult:
|
|
"""Send a structured generation request to Ollama and return the result.
|
|
|
|
Builds the Ollama-native payload, streams the response, applies
|
|
stall detection, and maps results to InferenceResult.
|
|
"""
|
|
start = time.monotonic()
|
|
|
|
payload = self._build_payload(request)
|
|
url = f"{self._target.base_url}/api/chat"
|
|
|
|
logger.info(
|
|
"Ollama POST %s model=%s messages=%d max_tokens=%d",
|
|
url,
|
|
self._target.model,
|
|
len(request.messages),
|
|
request.max_output_tokens,
|
|
)
|
|
|
|
try:
|
|
content, metadata = await self._stream_response(url, payload, request.timeout_seconds)
|
|
except InferenceError:
|
|
raise
|
|
except httpx.TimeoutException as exc:
|
|
raise InferenceError(
|
|
InferenceErrorCategory.TIMEOUT,
|
|
f"Request timed out after {request.timeout_seconds}s",
|
|
provider_detail=str(exc),
|
|
) from exc
|
|
except httpx.ConnectError as exc:
|
|
raise InferenceError(
|
|
InferenceErrorCategory.CONNECTION_REFUSED,
|
|
"Connection refused",
|
|
provider_detail=str(exc),
|
|
) from exc
|
|
except httpx.HTTPError as exc:
|
|
raise InferenceError(
|
|
InferenceErrorCategory.CONNECTION_ERROR,
|
|
"HTTP connection error",
|
|
provider_detail=str(exc),
|
|
) from exc
|
|
|
|
latency_ms = int((time.monotonic() - start) * 1000)
|
|
|
|
# Determine structured mode
|
|
structured_mode = self._determine_structured_mode(request)
|
|
|
|
# Attempt JSON parsing if we expect structured output
|
|
parsed = None
|
|
if request.json_schema and content:
|
|
try:
|
|
parsed = json.loads(content)
|
|
except (json.JSONDecodeError, ValueError):
|
|
pass
|
|
|
|
return InferenceResult(
|
|
content=content,
|
|
parsed=parsed,
|
|
endpoint_id=self._target.endpoint_id,
|
|
deployment_id=self._target.deployment_id,
|
|
model=metadata.get("model", self._target.model) or self._target.model,
|
|
protocol=self._target.protocol,
|
|
structured_mode=structured_mode,
|
|
latency_ms=latency_ms,
|
|
usage=TokenUsage(
|
|
input_tokens=metadata.get("prompt_eval_count"),
|
|
output_tokens=metadata.get("eval_count"),
|
|
),
|
|
request_id=request.trace_id or None,
|
|
repaired=False,
|
|
retries=0,
|
|
)
|
|
|
|
def _build_payload(self, request: StructuredGenerationRequest) -> dict:
|
|
"""Build the Ollama /api/chat request body."""
|
|
messages = [
|
|
{"role": msg.role, "content": msg.content}
|
|
for msg in request.messages
|
|
]
|
|
|
|
options: dict = {}
|
|
|
|
# Honor max output tokens via num_predict
|
|
if request.max_output_tokens:
|
|
options["num_predict"] = request.max_output_tokens
|
|
|
|
# Honor context window from target configuration
|
|
if self._target.context_window and self._target.context_window > 0:
|
|
options["num_ctx"] = self._target.context_window
|
|
|
|
# Temperature
|
|
options["temperature"] = request.temperature
|
|
|
|
# Seed if provided
|
|
if request.seed is not None:
|
|
options["seed"] = request.seed
|
|
|
|
payload: dict = {
|
|
"model": self._target.model,
|
|
"messages": messages,
|
|
"stream": True,
|
|
"options": options,
|
|
}
|
|
|
|
# Native schema formatting when supported
|
|
if request.json_schema and self._target.capabilities.json_schema:
|
|
payload["format"] = request.json_schema
|
|
|
|
# Merge any extra body params from target
|
|
if self._target.extra_body:
|
|
for key, value in self._target.extra_body.items():
|
|
if key not in payload:
|
|
payload[key] = value
|
|
|
|
return payload
|
|
|
|
def _determine_structured_mode(
|
|
self, request: StructuredGenerationRequest
|
|
) -> str:
|
|
"""Determine which structured output mode was used."""
|
|
if not request.json_schema:
|
|
return "none"
|
|
if self._target.capabilities.json_schema:
|
|
return "json_schema"
|
|
# Schema was requested but not natively supported — prompt only
|
|
return "prompt_only"
|
|
|
|
async def _stream_response(
|
|
self,
|
|
url: str,
|
|
payload: dict,
|
|
timeout_seconds: float,
|
|
) -> tuple[str, dict]:
|
|
"""Stream Ollama response, applying stall detection.
|
|
|
|
Returns (content, metadata) where metadata contains token counts
|
|
and timing from the final Ollama response chunk.
|
|
"""
|
|
content_parts: list[str] = []
|
|
metadata: dict = {}
|
|
last_content_length = 0
|
|
unchanged_count = 0
|
|
last_check_time = time.monotonic()
|
|
|
|
try:
|
|
async with self._http.stream(
|
|
"POST",
|
|
url,
|
|
json=payload,
|
|
timeout=httpx.Timeout(timeout_seconds, read=timeout_seconds),
|
|
) as response:
|
|
if response.status_code != 200:
|
|
# Read body for error detail
|
|
body = b""
|
|
async for chunk in response.aiter_bytes():
|
|
body += chunk
|
|
detail = body.decode("utf-8", errors="replace")[:500]
|
|
category = _map_http_status_to_category(response.status_code)
|
|
raise InferenceError(
|
|
category,
|
|
f"Ollama returned HTTP {response.status_code}",
|
|
provider_detail=detail,
|
|
status_code=response.status_code,
|
|
)
|
|
|
|
async for line in response.aiter_lines():
|
|
if not line.strip():
|
|
continue
|
|
|
|
try:
|
|
chunk_data = json.loads(line)
|
|
except json.JSONDecodeError:
|
|
continue
|
|
|
|
# Extract content from message
|
|
message = chunk_data.get("message", {})
|
|
chunk_content = message.get("content", "")
|
|
if chunk_content:
|
|
content_parts.append(chunk_content)
|
|
|
|
# Check for completion
|
|
if chunk_data.get("done", False):
|
|
# Capture metadata from final chunk
|
|
metadata["model"] = chunk_data.get("model", "")
|
|
metadata["prompt_eval_count"] = chunk_data.get("prompt_eval_count")
|
|
metadata["eval_count"] = chunk_data.get("eval_count")
|
|
total_ns = chunk_data.get("total_duration")
|
|
if total_ns:
|
|
metadata["total_duration_ms"] = total_ns // 1_000_000
|
|
break
|
|
|
|
# Stall detection
|
|
if self._stall_policy.enabled:
|
|
now = time.monotonic()
|
|
if now - last_check_time >= self._stall_policy.check_interval_seconds:
|
|
current_content = "".join(content_parts)
|
|
current_length = len(current_content)
|
|
|
|
if current_length == last_content_length:
|
|
unchanged_count += 1
|
|
else:
|
|
# Check for loop pattern
|
|
if _detect_loop(
|
|
current_content,
|
|
self._stall_policy.loop_window,
|
|
self._stall_policy.loop_threshold,
|
|
):
|
|
unchanged_count += 1
|
|
else:
|
|
unchanged_count = 0
|
|
|
|
last_content_length = current_length
|
|
last_check_time = now
|
|
|
|
if unchanged_count >= self._stall_policy.max_unchanged_intervals:
|
|
logger.warning(
|
|
"Stall detected after %d unchanged intervals, aborting",
|
|
unchanged_count,
|
|
)
|
|
raise InferenceError(
|
|
InferenceErrorCategory.STALL_DETECTED,
|
|
f"Generation stalled after {unchanged_count} check intervals",
|
|
)
|
|
|
|
except InferenceError:
|
|
raise
|
|
except httpx.TimeoutException as exc:
|
|
raise InferenceError(
|
|
InferenceErrorCategory.TIMEOUT,
|
|
"Stream read timed out",
|
|
provider_detail=str(exc),
|
|
) from exc
|
|
except httpx.ConnectError as exc:
|
|
raise InferenceError(
|
|
InferenceErrorCategory.CONNECTION_REFUSED,
|
|
"Connection refused during streaming",
|
|
provider_detail=str(exc),
|
|
) from exc
|
|
except httpx.HTTPError as exc:
|
|
raise InferenceError(
|
|
InferenceErrorCategory.CONNECTION_ERROR,
|
|
"HTTP error during streaming",
|
|
provider_detail=str(exc),
|
|
) from exc
|
|
|
|
final_content = "".join(content_parts)
|
|
|
|
if not final_content:
|
|
raise InferenceError(
|
|
InferenceErrorCategory.EMPTY_RESPONSE,
|
|
"Ollama returned empty content",
|
|
)
|
|
|
|
return final_content, metadata
|