E2e workflow (#21)

Reviewed-on: #21
Co-authored-by: Mathias Scheider <mathias2784@web.de>

.env.template

@@ -3,3 +3,8 @@ LLM_API_KEY = "not-relevant"
 LLM_DEFAULT_TEMPERATURE = 1.0
 
 METADATA_PREFIX = "photools|"
+LOG_LEVEL=INFO
+LLM_REQUEST_TIMEOUT_SECONDS=10
+LLM_MAX_COMPLETION_TOKENS=1024
+LLM_MAX_RETRIES=5
+LLM_MAX_KEYWORDS=15

.gitea/workflows/ci.yml

@@ -35,5 +35,3 @@ jobs:
         run: uv run mypy .
       - name: Check formatting (black)
         run: uv run black --check . --extend-exclude .venv
-      - name: Check formatting (black)
-        run: uv run black --check . --extend-exclude .venv

AGENTS.md

@@ -1,27 +1,42 @@
 # Agent Instructions: PhotoTools3
 
 ## Environment Setup
+
 Requires a `.env` file with the following variables:
+
 - `llm_api_key`: API key for the LLM provider.
 - `llm_api_base_url`: Base URL for the LLM API.
 - `llm_default_temperature`: Float value for temperature.
 
 ## System Dependencies
+
 - `exiftool` binary: required by `pyexiftool` at runtime, not installable via pip/uv.
   - Arch: `perl-image-exiftool`
   - Debian/Ubuntu: `libimage-exiftool-perl`
 
 ## Development Toolchain
+
 - **Dependency Management**: Use `uv`.
 - **Python Version**: Requires Python `>=3.14`.
 - **Testing**: Run `pytest` to execute the test suite in `test/`.
 - **Linting & Formatting**: Use `ruff` and `black`.
 
 ## CLI Usage
+
 The primary entry point is the `photools3` command:
+
 - `photools3 keywords <image_path>`: Extracts keywords and captions from a single image.
 
 ## Architecture & Logic
+
 - **`src/photools3/photo.py`**: Core image processing logic, including resizing, RGB conversion, and generating base64 JPEG encodings specifically for LLM vision compatibility.
 - **`src/photools3/llm_session.py`**: Manages the workflow of sending processed images to the LLM and parsing results.
 - **Configuration**: The list of keywords to extract is maintained in `config/keywords.txt`.
+
+## Code style
+
+- Avoid overly verbose docstrings or comments. Be precise!
+
+## Comunication
+
+- When done with code modifications, give a one-liner for the currently unstaged changes that can be used as a commit message.

README.md

@@ -17,13 +17,28 @@ Create a `.env` file in the project root with the LLM connection details and the
 prefix used to identify metadata written by PhotoTools3:
 
 ```dotenv
-llm_api_base_url=https://your-openai-compatible-api/v1
+llm_api_base_url=http://localhost:1234/v1
 llm_api_key=your-api-key
 llm_default_temperature=0.2
 metadata_prefix=photools3:
+LOG_LEVEL=DEBUG
+LLM_REQUEST_TIMEOUT_SECONDS=10
+LLM_MAX_COMPLETION_TOKENS=1024
+LLM_MAX_RETRIES=5
+LLM_MAX_KEYWORDS=15
 ```
 
 Keyword suggestions are restricted to the vocabulary in `config/keywords.txt`.
+`LOG_LEVEL=DEBUG` enables detailed per-request and per-photo timings.
+`LLM_MAX_RETRIES` controls retries for malformed or structurally invalid
+responses; `5` means up to 6 requests for a photo. A valid response with
+unsupported keywords gets at most one targeted correction retry, then remaining
+invalid keywords are discarded.
+Each request is limited by `LLM_REQUEST_TIMEOUT_SECONDS`; a timeout reloads the
+LM Studio model before retrying. `LLM_MAX_COMPLETION_TOKENS` caps the generated
+response length. When all retries are exhausted, the model is reloaded before
+the failed photo is reported. `LLM_MAX_KEYWORDS` sets the maximum number of
+distinct keywords requested and stored per photo; excess keywords are discarded.
 
 ## Generate metadata
 

pyproject.toml

@@ -18,7 +18,7 @@ dependencies = [
     "pytest>=9.1.1",
     "ruff>=0.16.9",
     "typer>=0.12.0",
-    "utils>=0.1.2",
+    "utils>=0.1.3",
 ]
 
 [project.scripts]

src/photools3/cli.py

@@ -13,9 +13,20 @@ from photools3.processing import (
     process_batch,
     process_photo,
 )
+from photools3.settings import settings
 
 app = typer.Typer(help="PhotoTools3: A CLI for image processing.")
-LoggingSettings()
+LoggingSettings(LOG_LEVEL=settings.log_level)
+for name in (
+    "uvicorn",
+    "uvicorn.error",
+    "uvicorn.access",
+    "PIL",
+    "openai",
+    "httpx2",
+    "httpcore2",
+):
+    logging.getLogger(name).setLevel(logging.WARNING)
 logger = logging.getLogger(__name__)
 
 

src/photools3/llm_session.py

@@ -1,11 +1,13 @@
 import json
 import logging
 import re
+import time
 
 import openai
 from openai import OpenAI
 from pydantic import BaseModel, ValidationError
 
+from photools3.lm_studio import LMStudioModelManager
 from photools3.models import (
     LlmMessageContent,
     LlmMessageRole,
@@ -17,6 +19,7 @@ from photools3.models import (
     LlmResponseGeneral,
     LlmResponseKeywords,
     LlmResponseQuality,
+    load_valid_keywords,
 )
 from photools3.photo import Photo
 from photools3.settings import settings
@@ -30,19 +33,23 @@ def _clean_llm_response(txt: str):
     return txt
 
 
-def _correction_prompt(
-    original_request: LlmRequest, exception: Exception, origial_json
-):
-    correction_prompt = (
-        f"The previous response was invalid due to: {exception}. "
-        f"Please provide the corrected JSON based on this content: "
-        f"{origial_json}"
+def _correction_prompt(instruction: str) -> LlmRequestMessage:
+    return LlmRequestMessage(
+        role=LlmMessageRole.user,
+        content=instruction,
     )
-    original_request.messages.append(
-        LlmRequestMessage(
-            role=LlmMessageRole.user,
-            content=correction_prompt,
-        )
+
+
+def _request_with_correction(base_request: LlmRequest, instruction: str) -> LlmRequest:
+    return base_request.model_copy(
+        update={"messages": [*base_request.messages, _correction_prompt(instruction)]}
+    )
+
+
+def _validation_summary(exception: ValidationError) -> str:
+    return "; ".join(
+        f"{'.'.join(map(str, error['loc']))}: {error['msg']}"
+        for error in exception.errors(include_input=False)
     )
 
 
@@ -56,6 +63,7 @@ class LlmSession:
         self, system_prompt: str, temperature: float = settings.llm_default_temperature
     ):
         self._models: list = []
+        self._model_id: str | None = None
 
         self.system_prompt = system_prompt
         self.temperature = temperature
@@ -63,7 +71,9 @@ class LlmSession:
         self.client = OpenAI(
             base_url=self.base_url,
             api_key=self.api_key,
+            max_retries=0,
         )
+        self._model_manager = LMStudioModelManager(self.base_url, self.api_key)
 
     def close(self):
         self.client.close()
@@ -74,8 +84,17 @@ class LlmSession:
     def __exit__(self, exc_type, exc_value, traceback):
         self.close()
 
+    def flush_model(self):
+        """Unload and reload this session's model through LM Studio."""
+        if self._model_id is None:
+            self._model_id = self.models[0].id
+        self._model_manager.reload_model(self._model_id)
+
     def client_chat_completion_create(self, request: LlmRequest, choice_no: int = 0):
-        response = self.client.chat.completions.create(**request.model_dump())
+        response = self.client.chat.completions.create(
+            **request.model_dump(),
+            timeout=settings.llm_request_timeout_seconds,
+        )
         return response.choices[choice_no].message.content
 
     def is_api_reachable(self) -> bool:
@@ -95,37 +114,139 @@ class LlmSession:
         self,
         photos: list[Photo],
         model_id=None,
-        max_retries: int = 2,
+        max_retries: int | None = None,
         prompt: str | None = None,
     ):
-        if not model_id:
-            model_id = self.models[0].id
+        retry_limit = settings.llm_max_retries if max_retries is None else max_retries
+        if retry_limit < 0:
+            raise ValueError("max_retries must be greater than or equal to zero")
 
-        request = self._build_request(prompt or self.prompt, photos, model_id)
+        if model_id is None:
+            if self._model_id is None:
+                self._model_id = self.models[0].id
+            model_id = self._model_id
+        else:
+            self._model_id = model_id
 
-        last_error = None
-        for attempt in range(max_retries + 1):
+        base_request = self._build_request(prompt or self.prompt, photos, model_id)
+        request = base_request
+        tag_correction_used = False
+        valid_keywords = load_valid_keywords()
+        max_keywords = settings.llm_max_keywords
+
+        for attempt in range(retry_limit + 1):
+            request_started = time.perf_counter()
             try:
                 result_raw = self.client_chat_completion_create(request)
-                json_str = _clean_llm_response(result_raw)
-                return self.response_model.model_validate_json(json_str)
+            except openai.APITimeoutError as exc:
+                logger.warning(
+                    "LLM request timed out after %.2f seconds; reloading model",
+                    time.perf_counter() - request_started,
+                )
+                self.flush_model()
+                if attempt < retry_limit:
+                    continue
+                raise ValueError(
+                    f"LLM request timed out after {retry_limit + 1} attempts"
+                ) from exc
+            except Exception:
+                logger.debug(
+                    "LLM request attempt %d/%d failed after %.2f seconds",
+                    attempt + 1,
+                    retry_limit + 1,
+                    time.perf_counter() - request_started,
+                )
+                raise
 
-            except (json.JSONDecodeError, ValidationError) as exc:
-                last_error = exc
-                if attempt < max_retries:
-                    logger.info("Chatbot response is invalid; promting to correct ...")
-                    _correction_prompt(
-                        request,
-                        exc,
-                        json_str if "json_str" in locals() else "origial json",
+            logger.debug(
+                "LLM request attempt %d/%d completed in %.2f seconds",
+                attempt + 1,
+                retry_limit + 1,
+                time.perf_counter() - request_started,
+            )
+            try:
+                response_data = json.loads(_clean_llm_response(result_raw))
+            except (json.JSONDecodeError, TypeError) as exc:
+                if attempt < retry_limit:
+                    logger.info("Chatbot response was not valid JSON; retrying")
+                    request = _request_with_correction(
+                        base_request,
+                        "Return one complete JSON object matching the schema. "
+                        f"The previous response was invalid: {exc}",
                     )
                     continue
-                else:
-                    raise ValueError(
-                        f"Failed to parse or validate LLM response after {max_retries} retries: {exc}"
+
+                logger.warning(
+                    "LLM retry limit reached; reloading model before failing"
+                )
+                self.flush_model()
+                raise ValueError(
+                    f"Failed to parse LLM response after {retry_limit} retries: {exc}"
+                ) from exc
+
+            if (
+                isinstance(response_data, dict)
+                and "tags" in self.response_model.model_fields
+                and isinstance(response_data.get("tags"), list)
+            ):
+                tags = response_data["tags"]
+                considered_tags = tags[:max_keywords]
+                invalid_tags = [
+                    tag
+                    for tag in considered_tags
+                    if not isinstance(tag, str) or tag not in valid_keywords
+                ]
+
+                if invalid_tags and not tag_correction_used and attempt < retry_limit:
+                    tag_correction_used = True
+                    logger.info(
+                        "LLM returned unsupported keywords; requesting one correction"
                     )
-            except Exception:
-                raise
+                    request = _request_with_correction(
+                        base_request,
+                        f"These keywords are not allowed: {invalid_tags!r}. "
+                        "Replace them with exact allowed keywords when sensible; "
+                        f"otherwise omit them. Return at most {max_keywords} "
+                        "distinct keywords and do not list the vocabulary.",
+                    )
+                    continue
+
+                if len(tags) > max_keywords:
+                    logger.info(
+                        "Discarding keywords beyond configured limit of %d",
+                        max_keywords,
+                    )
+                if invalid_tags:
+                    logger.warning("Discarding unsupported keywords: %s", invalid_tags)
+
+                response_data = dict(response_data)
+                valid_tags = [
+                    tag
+                    for tag in tags
+                    if isinstance(tag, str) and tag in valid_keywords
+                ]
+                response_data["tags"] = list(dict.fromkeys(valid_tags))[:max_keywords]
+
+            try:
+                return self.response_model.model_validate(response_data)
+            except ValidationError as exc:
+                if attempt < retry_limit:
+                    logger.info("LLM response did not match the schema; retrying")
+                    request = _request_with_correction(
+                        base_request,
+                        "Return one complete JSON object matching the schema. "
+                        f"Validation issues: {_validation_summary(exc)}",
+                    )
+                    continue
+
+                logger.warning(
+                    "LLM retry limit reached; reloading model before failing"
+                )
+                self.flush_model()
+                raise ValueError(
+                    f"Failed to validate LLM response after {retry_limit} retries: "
+                    f"{_validation_summary(exc)}"
+                ) from exc
 
     def _build_request(
         self,
@@ -137,6 +258,7 @@ class LlmSession:
         return LlmRequest(
             model=model_id,
             temperature=self.temperature,
+            max_tokens=settings.llm_max_completion_tokens,
             messages=[
                 # System prompt
                 LlmRequestMessage(
@@ -168,13 +290,15 @@ class LlmSession:
 
 class LlmSessionCaptionsAndKeywords(LlmSession):
     response_model = LlmResponseCaptionsAndKeywords
-    prompt = "Give me a caption and keyword for the following photo(s)"
+    prompt = "Give me a caption and keywords for the following photo"
 
     def __init__(self, temperature: float = settings.llm_default_temperature):
         super().__init__(
             system_prompt=(
                 "You are a helpful photographer's assistant that helps to write **captions** and **keywords** to photos:\n"
                 "Answer in German.\n"
+                f"Choose at most {settings.llm_max_keywords} distinct, relevant keywords. "
+                "Do not repeat or enumerate the allowed-keyword list.\n"
                 "Output must be a JSON of the following schema:\n"
                 f"{self.response_model.model_json_schema()}"
             ),
@@ -207,6 +331,8 @@ class LlmSessionKeywords(LlmSession):
             system_prompt=(
                 "You are a helpful photographer's assistant that chooses keywords "
                 "for photos. Answer in German.\n"
+                f"Choose at most {settings.llm_max_keywords} distinct, relevant keywords. "
+                "Do not repeat or enumerate the allowed-keyword list.\n"
                 "Output must be a JSON of the following schema:\n"
                 f"{self.response_model.model_json_schema()}"
             ),

src/photools3/lm_studio.py

@@ -0,0 +1,135 @@
+import json
+import logging
+import time
+from urllib.error import HTTPError, URLError
+from urllib.parse import urlsplit, urlunsplit
+from urllib.request import Request, urlopen
+
+logger = logging.getLogger(__name__)
+
+_LOAD_CONFIG_KEYS = (
+    "context_length",
+    "eval_batch_size",
+    "flash_attention",
+    "num_experts",
+    "offload_kv_cache_to_gpu",
+)
+
+
+class LMStudioError(RuntimeError):
+    """Raised when LM Studio model management fails."""
+
+
+class LMStudioModelManager:
+    """Unload and reload a model through LM Studio's native REST API."""
+
+    def __init__(self, openai_base_url: str, api_key: str, timeout: float = 600):
+        parsed_url = urlsplit(openai_base_url)
+        path = parsed_url.path.rstrip("/")
+        path = path.removesuffix("/v1")
+        root_url = urlunsplit(
+            (parsed_url.scheme, parsed_url.netloc, path, "", "")
+        ).rstrip("/")
+        self.api_base_url = f"{root_url}/api/v1"
+        self.api_key = api_key
+        self.timeout = timeout
+
+    def reload_model(self, model_id: str) -> None:
+        """Unload the matching loaded instance and reload it with its config."""
+        started = time.perf_counter()
+        models_response = self._request("GET", "/models")
+        model, loaded_instance = self._find_model(models_response, model_id)
+        model_key = model["key"]
+
+        if loaded_instance is not None:
+            self._request(
+                "POST",
+                "/models/unload",
+                {"instance_id": loaded_instance["id"]},
+            )
+            load_config = loaded_instance.get("config", {})
+        else:
+            load_config = {}
+
+        load_body = {"model": model_key}
+        load_body.update(
+            {
+                key: load_config[key]
+                for key in _LOAD_CONFIG_KEYS
+                if key in load_config and load_config[key] is not None
+            }
+        )
+        self._request("POST", "/models/load", load_body)
+        logger.info(
+            "Reloaded LM Studio model %s in %.1f seconds",
+            model_key,
+            time.perf_counter() - started,
+        )
+
+    @staticmethod
+    def _find_model(models_response: dict, model_id: str) -> tuple[dict, dict | None]:
+        models = models_response.get("models")
+        if not isinstance(models, list):
+            raise LMStudioError("LM Studio model list response has no models array")
+
+        for model in models:
+            loaded_instances = model.get("loaded_instances", [])
+            for instance in loaded_instances:
+                if instance.get("id") == model_id:
+                    return model, instance
+
+        matching_models = [model for model in models if model.get("key") == model_id]
+        if len(matching_models) != 1:
+            raise LMStudioError(
+                f"Could not uniquely identify LM Studio model {model_id!r}"
+            )
+
+        model = matching_models[0]
+        loaded_instances = model.get("loaded_instances", [])
+        if len(loaded_instances) > 1:
+            raise LMStudioError(
+                f"Multiple loaded instances found for LM Studio model {model_id!r}"
+            )
+        loaded_instance = loaded_instances[0] if loaded_instances else None
+        return model, loaded_instance
+
+    def _request(self, method: str, path: str, body: dict | None = None) -> dict:
+        data = None if body is None else json.dumps(body).encode("utf-8")
+        headers = {"Accept": "application/json"}
+        if data is not None:
+            headers["Content-Type"] = "application/json"
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+
+        request = Request(
+            f"{self.api_base_url}{path}",
+            data=data,
+            headers=headers,
+            method=method,
+        )
+        try:
+            with urlopen(request, timeout=self.timeout) as response:
+                response_body = response.read()
+        except HTTPError as exc:
+            detail = exc.read().decode("utf-8", errors="replace")
+            raise LMStudioError(
+                f"LM Studio {method} {path} returned HTTP {exc.code}: {detail}"
+            ) from exc
+        except (URLError, TimeoutError) as exc:
+            raise LMStudioError(
+                f"Could not reach LM Studio at {self.api_base_url}: {exc}"
+            ) from exc
+
+        if not response_body:
+            return {}
+        try:
+            result = json.loads(response_body)
+        except json.JSONDecodeError as exc:
+            raise LMStudioError(
+                f"LM Studio {method} {path} returned invalid JSON"
+            ) from exc
+        if not isinstance(result, dict):
+            raise LMStudioError(
+                f"LM Studio {method} {path} returned an unexpected response"
+            )
+        return result

src/photools3/models.py

@@ -1,3 +1,4 @@
+import json
 from enum import StrEnum
 from functools import lru_cache
 from typing import Annotated, Any, Literal
@@ -10,7 +11,7 @@ from pydantic import (
     model_validator,
 )
 
-from .settings import KEYWORDS_LIST_PATH
+from .settings import KEYWORDS_LIST_PATH, settings
 
 
 @lru_cache
@@ -42,6 +43,7 @@ class LlmRequest(BaseModel):
     model: str = "gpt-4o"
     messages: list[LlmRequestMessage]
     temperature: float | None = None
+    max_tokens: int
 
 
 class LlmMessageRole(StrEnum):
@@ -98,17 +100,22 @@ class LlmResponseCaption(BaseModel):
     ]
 
 
-class LlmResponseKeywords(BaseModel):
-    tags: list[str] = Field(
-        description=(
-            "A list of keywords that properly tag the photo. "
-            "Allowed keywords: "
-            + str(load_valid_keywords())
-            .replace("'", '"')
-            .replace("{", "[")
-            .replace("}", "]")
-        )
+def _keyword_tags_field():
+    allowed_keywords = json.dumps(
+        sorted(load_valid_keywords()), ensure_ascii=False, separators=(",", ":")
     )
+    return Field(
+        max_length=settings.llm_max_keywords,
+        description=(
+            f"Choose at most {settings.llm_max_keywords} distinct, relevant keywords. "
+            "Do not enumerate the vocabulary. Allowed keywords: "
+            f"{allowed_keywords}"
+        ),
+    )
+
+
+class _LlmResponseWithTags(BaseModel):
+    tags: list[str] = _keyword_tags_field()
 
     @field_validator("tags")
     @classmethod
@@ -122,7 +129,11 @@ class LlmResponseKeywords(BaseModel):
         return v
 
 
-class LlmResponseCaptionsAndKeywords(BaseModel):
+class LlmResponseKeywords(_LlmResponseWithTags):
+    pass
+
+
+class LlmResponseCaptionsAndKeywords(_LlmResponseWithTags):
     caption: Annotated[
         str,
         Field(
@@ -132,27 +143,6 @@ class LlmResponseCaptionsAndKeywords(BaseModel):
             )
         ),
     ]
-    tags: list[str] = Field(
-        description=(
-            "A list of keywords that properly tag the photo. "
-            "Allowed keywords: "
-            + str(load_valid_keywords())
-            .replace("'", '"')
-            .replace("{", "[")
-            .replace("}", "]")
-        )
-    )
-
-    @field_validator("tags")
-    @classmethod
-    def validate_tags(cls, v: list[str]) -> list[str]:
-        valid_keywords = load_valid_keywords()
-        for tag in v:
-            if tag not in valid_keywords:
-                raise ValueError(
-                    f"'{tag}' is not a valid keyword. Must be one from the configuration."
-                )
-        return v
 
 
 class LlmResponseQuality(BaseModel):

src/photools3/processing.py

@@ -1,4 +1,5 @@
 import logging
+import time
 from collections.abc import Iterable
 from dataclasses import dataclass
 from enum import StrEnum
@@ -65,6 +66,8 @@ def process_photo(
     owns_session = session is None
     active_session = session if session is not None else _create_session(mode)
 
+    logger.info(f"Processing photo: {path!s}")
+
     try:
         with Photo(path) as photo:
             result = active_session.call([photo])
@@ -108,6 +111,7 @@ def process_batch(
     try:
         for item in paths:
             path = Path(item)
+            photo_started = time.perf_counter()
             try:
                 result = process_photo(path, mode, session=session)
             except Exception as exc:
@@ -122,6 +126,12 @@ def process_batch(
                 )
             else:
                 outcomes.append(ProcessingOutcome(path=path, result=result))
+            finally:
+                logger.debug(
+                    "Photo processing for %s took %.2f seconds",
+                    path,
+                    time.perf_counter() - photo_started,
+                )
     finally:
         session.close()
 

src/photools3/settings.py

@@ -1,5 +1,6 @@
 from pathlib import Path
 
+from pydantic import Field
 from pydantic_core import Url
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
@@ -13,11 +14,16 @@ KEYWORDS_LIST_PATH = CONFIG_DIR / "keywords.txt"
 
 
 class Settings(BaseSettings):
-    model_config = SettingsConfigDict(env_file=".env")
+    model_config = SettingsConfigDict(env_file=".env", extra="ignore")
 
     llm_api_base_url: Url
     llm_api_key: str
     llm_default_temperature: float
+    llm_request_timeout_seconds: float = Field(default=10, gt=0)
+    llm_max_completion_tokens: int = Field(default=1024, gt=0)
+    llm_max_retries: int = Field(default=5, ge=0)
+    llm_max_keywords: int = Field(default=15, ge=1)
+    log_level: str = "INFO"
 
     metadata_prefix: str
 

test/test_llm_session_mocked.py

@@ -1,6 +1,10 @@
 from unittest.mock import MagicMock, patch
 
-from photools3.llm_session import LlmSession
+import openai
+import pytest
+
+from photools3.llm_session import LlmSession, LlmSessionCaptionsAndKeywords
+from photools3.models import LlmResponseCaptionsAndKeywords
 
 
 def test_llm_session_mocked(test_photo):
@@ -84,3 +88,216 @@ def test_llm_session_captions_and_keywords_mocked(test_photo):
 
         # Verify that the mock was called
         assert mock_client.chat.completions.create.called
+
+
+def test_llm_session_retries_from_original_request_without_accumulating_messages():
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        def response(content):
+            message = MagicMock()
+            message.content = content
+            choice = MagicMock()
+            choice.message = message
+            result = MagicMock()
+            result.choices = [choice]
+            return result
+
+        mock_client.chat.completions.create.side_effect = [
+            response('{"wrong": "first"}'),
+            response('{"wrong": "second"}'),
+            response('{"text": "correct"}'),
+        ]
+
+        session = LlmSession(system_prompt="You are a helper.")
+        result = session.call(photos=[], prompt="Return JSON")
+
+    assert result.text == "correct"
+    requests = mock_client.chat.completions.create.call_args_list
+    messages_per_attempt = [call.kwargs["messages"] for call in requests]
+    assert [len(messages) for messages in messages_per_attempt] == [2, 3, 3]
+    final_correction = messages_per_attempt[2][-1]["content"]
+    assert "Field required" in final_correction
+    assert '"wrong"' not in final_correction
+
+
+def test_llm_session_uses_configured_retry_count(monkeypatch):
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_retries", 1)
+
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        def response(content):
+            message = MagicMock()
+            message.content = content
+            choice = MagicMock()
+            choice.message = message
+            result = MagicMock()
+            result.choices = [choice]
+            return result
+
+        mock_client.chat.completions.create.side_effect = [
+            response('{"wrong": "first"}'),
+            response('{"wrong": "second"}'),
+            response('{"text": "would be a third call"}'),
+        ]
+        session = LlmSession(system_prompt="You are a helper.")
+
+        with (
+            patch.object(session, "flush_model") as flush_model,
+            pytest.raises(ValueError, match="after 1 retries"),
+        ):
+            session.call(photos=[])
+
+    assert mock_client.chat.completions.create.call_count == 2
+    flush_model.assert_called_once()
+
+
+def test_llm_session_reloads_after_timeout_and_retries(monkeypatch):
+    monkeypatch.setattr(
+        "photools3.llm_session.settings.llm_request_timeout_seconds", 10
+    )
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_completion_tokens", 768)
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_retries", 1)
+
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        response_message = MagicMock(content='{"text": "recovered"}')
+        response_choice = MagicMock(message=response_message)
+        success_response = MagicMock(choices=[response_choice])
+        mock_client.chat.completions.create.side_effect = [
+            openai.APITimeoutError(request=MagicMock()),
+            success_response,
+        ]
+
+        session = LlmSession(system_prompt="You are a helper.")
+        with patch.object(session, "flush_model") as flush_model:
+            result = session.call(photos=[])
+
+    assert result.text == "recovered"
+    flush_model.assert_called_once()
+    calls = mock_client.chat.completions.create.call_args_list
+    assert len(calls) == 2
+    assert all(call.kwargs["timeout"] == 10 for call in calls)
+    assert all(call.kwargs["max_tokens"] == 768 for call in calls)
+    mocked_openai.assert_called_once_with(
+        base_url=session.base_url,
+        api_key=session.api_key,
+        max_retries=0,
+    )
+
+
+@pytest.mark.parametrize("failure_kind", ["invalid_response", "timeout"])
+def test_llm_session_reloads_when_retry_limit_is_exhausted(monkeypatch, failure_kind):
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_retries", 0)
+
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        if failure_kind == "timeout":
+            mock_client.chat.completions.create.side_effect = openai.APITimeoutError(
+                request=MagicMock()
+            )
+        else:
+            response_message = MagicMock(content='{"wrong": "shape"}')
+            response_choice = MagicMock(message=response_message)
+            mock_client.chat.completions.create.return_value = MagicMock(
+                choices=[response_choice]
+            )
+
+        session = LlmSession(system_prompt="You are a helper.")
+        with (
+            patch.object(session, "flush_model") as flush_model,
+            pytest.raises(ValueError),
+        ):
+            session.call(photos=[])
+
+    flush_model.assert_called_once()
+    assert mock_client.chat.completions.create.call_count == 1
+
+
+def test_invalid_keywords_get_one_correction_then_invalid_tags_are_dropped(monkeypatch):
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_keywords", 3)
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_retries", 5)
+
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        def response(content):
+            message = MagicMock(content=content)
+            choice = MagicMock(message=message)
+            return MagicMock(choices=[choice])
+
+        mock_client.chat.completions.create.side_effect = [
+            response(
+                '{"caption": "A caption", "tags": ["not_allowed", "architektur"]}'
+            ),
+            response(
+                '{"caption": "A caption", '
+                '"tags": ["still_invalid", "architektur", "personen"]}'
+            ),
+        ]
+        session = LlmSessionCaptionsAndKeywords()
+        with patch.object(session, "flush_model") as flush_model:
+            result = session.call(photos=[])
+
+    assert result.tags == ["architektur", "personen"]
+    assert mock_client.chat.completions.create.call_count == 2
+    flush_model.assert_not_called()
+    correction = mock_client.chat.completions.create.call_args_list[1].kwargs[
+        "messages"
+    ][-1]["content"]
+    assert "not_allowed" in correction
+    assert '"caption"' not in correction
+
+
+def test_excess_keywords_are_truncated_to_configured_limit(monkeypatch):
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_keywords", 2)
+
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+        message = MagicMock(
+            content=(
+                '{"caption": "A caption", '
+                '"tags": ["architektur", "personen", "landschaft"]}'
+            )
+        )
+        mock_client.chat.completions.create.return_value = MagicMock(
+            choices=[MagicMock(message=message)]
+        )
+
+        result = LlmSessionCaptionsAndKeywords().call(photos=[])
+
+    assert result.tags == ["architektur", "personen"]
+    assert mock_client.chat.completions.create.call_count == 1
+
+
+def test_keyword_response_schema_includes_configured_max_items():
+    schema = LlmResponseCaptionsAndKeywords.model_json_schema()
+
+    assert schema["properties"]["tags"]["maxItems"] == 15

test/test_lm_studio.py

@@ -0,0 +1,102 @@
+import json
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from photools3.lm_studio import LMStudioError, LMStudioModelManager
+
+
+def _http_response(payload: dict):
+    response = MagicMock()
+    response.__enter__.return_value = response
+    response.read.return_value = json.dumps(payload).encode("utf-8")
+    return response
+
+
+def test_reload_model_unloads_and_reloads_preserving_context_config():
+    model_response = {
+        "models": [
+            {
+                "key": "qwen/qwen3-vl-4b",
+                "loaded_instances": [
+                    {
+                        "id": "qwen/qwen3-vl-4b-instance",
+                        "config": {
+                            "context_length": 16384,
+                            "eval_batch_size": 256,
+                            "flash_attention": True,
+                            "offload_kv_cache_to_gpu": False,
+                            "parallel": 2,
+                        },
+                    }
+                ],
+            }
+        ]
+    }
+    with patch(
+        "photools3.lm_studio.urlopen",
+        side_effect=[
+            _http_response(model_response),
+            _http_response({"instance_id": "qwen/qwen3-vl-4b-instance"}),
+            _http_response({"instance_id": "qwen/qwen3-vl-4b-instance"}),
+        ],
+    ) as mocked_urlopen:
+        manager = LMStudioModelManager(
+            "http://localhost:1234/v1", "test-token", timeout=12
+        )
+        manager.reload_model("qwen/qwen3-vl-4b")
+
+    requests = [call.args[0] for call in mocked_urlopen.call_args_list]
+    assert requests[0].full_url == "http://localhost:1234/api/v1/models"
+    assert requests[0].get_method() == "GET"
+    assert requests[0].get_header("Authorization") == "Bearer test-token"
+
+    assert requests[1].full_url.endswith("/models/unload")
+    assert json.loads(requests[1].data) == {"instance_id": "qwen/qwen3-vl-4b-instance"}
+
+    assert requests[2].full_url.endswith("/models/load")
+    assert json.loads(requests[2].data) == {
+        "model": "qwen/qwen3-vl-4b",
+        "context_length": 16384,
+        "eval_batch_size": 256,
+        "flash_attention": True,
+        "offload_kv_cache_to_gpu": False,
+    }
+
+
+def test_reload_model_loads_by_key_if_no_instance_is_loaded():
+    model_response = {"models": [{"key": "qwen/model", "loaded_instances": []}]}
+    with patch(
+        "photools3.lm_studio.urlopen",
+        side_effect=[
+            _http_response(model_response),
+            _http_response({"status": "loaded"}),
+        ],
+    ) as mocked_urlopen:
+        manager = LMStudioModelManager("http://localhost:1234/v1", "token")
+        manager.reload_model("qwen/model")
+
+    assert mocked_urlopen.call_count == 2
+    load_request = mocked_urlopen.call_args.args[0]
+    assert load_request.full_url.endswith("/models/load")
+    assert json.loads(load_request.data) == {"model": "qwen/model"}
+
+
+def test_reload_model_rejects_ambiguous_model_selection():
+    model_response = {
+        "models": [
+            {
+                "key": "qwen/model",
+                "loaded_instances": [
+                    {"id": "instance-1", "config": {}},
+                    {"id": "instance-2", "config": {}},
+                ],
+            }
+        ]
+    }
+    with patch(
+        "photools3.lm_studio.urlopen", return_value=_http_response(model_response)
+    ):
+        manager = LMStudioModelManager("http://localhost:1234/v1", "token")
+        with pytest.raises(LMStudioError, match="Multiple loaded instances"):
+            manager.reload_model("qwen/model")

uv.lock

@@ -780,7 +780,7 @@ requires-dist = [
     { name = "pytest", specifier = ">=9.1.1" },
     { name = "ruff", specifier = ">=0.16.9" },
     { name = "typer", specifier = ">=0.12.0" },
-    { name = "utils", specifier = ">=0.1.2", index = "https://gitea.schneiderserver.site/api/packages/mathias2784/pypi/simple" },
+    { name = "utils", specifier = ">=0.1.3", index = "https://gitea.schneiderserver.site/api/packages/mathias2784/pypi/simple" },
 ]
 
 [[package]]
@@ -1205,7 +1205,7 @@ wheels = [
 
 [[package]]
 name = "utils"
-version = "0.1.2"
+version = "0.1.3"
 source = { registry = "https://gitea.schneiderserver.site/api/packages/mathias2784/pypi/simple" }
 dependencies = [
     { name = "black" },
@@ -1217,9 +1217,9 @@ dependencies = [
     { name = "twine" },
     { name = "uvicorn" },
 ]
-sdist = { url = "https://gitea.schneiderserver.site/api/packages/mathias2784/pypi/files/utils/0.1.2/utils-0.1.2.tar.gz", hash = "sha256:c711755c3e2fabd9cd380e25e45bc19da9b901868a6eb67028968803c725b118" }
+sdist = { url = "https://gitea.schneiderserver.site/api/packages/mathias2784/pypi/files/utils/0.1.3/utils-0.1.3.tar.gz", hash = "sha256:8970a179573a63ab4c546cfd1da8ceca0d915bc4357b49ccffaba9b3d4a0195e" }
 wheels = [
-    { url = "https://gitea.schneiderserver.site/api/packages/mathias2784/pypi/files/utils/0.1.2/utils-0.1.2-py3-none-any.whl", hash = "sha256:f3335ba10bd8eb9f5c000287a6ee41987b188372f3dca2783d99e6e70513a5d7" },
+    { url = "https://gitea.schneiderserver.site/api/packages/mathias2784/pypi/files/utils/0.1.3/utils-0.1.3-py3-none-any.whl", hash = "sha256:c24078ba05f9f5b41fc3909598084b93c6df59f4ca60001e1e0b4fa55797e642" },
 ]
 
 [[package]]