Cap generated keywords and simplify LLM correction retries.

README.md

@@ -22,18 +22,23 @@ llm_api_key=your-api-key
 llm_default_temperature=0.2
 metadata_prefix=photools3:
 LOG_LEVEL=DEBUG
-LLM_FLUSH_EVERY_PHOTOS=50
+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. The flush
-interval uses LM Studio's native model-management API to unload and reload the
-model after every 50 attempted batch photos. This requires LM Studio's native
-`/api/v1` model-management API. Set `LLM_FLUSH_EVERY_PHOTOS=0` to disable it;
-use `0` for non-LM-Studio OpenAI-compatible providers.
-`LLM_MAX_RETRIES` controls how many corrective retry requests are allowed after
-the initial LLM response; `5` means up to 6 requests for a photo.
+`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