Claus Ibsen created CAMEL-26219:
-----------------------------------
Summary: camel-huggingface - the chat history map grows without
bound, and any output containing the text "error" is treated as a failure
Key: CAMEL-26219
URL: https://issues.apache.org/jira/browse/CAMEL-26219
Project: Camel
Issue Type: Bug
Components: camel-ai
Reporter: Claus Ibsen
Fix For: 4.24.0
Found while adding the catalog body types of huggingface (CAMEL-24961, PR
#28402).
h4. 1. The chat history map grows without bound
{{ChatPredictor}} (main 2a465aa4814c, line 103) keeps a {{ConcurrentHashMap}}
of histories keyed by the memory id header (line 122); an entry is removed only
with {{CamelChatClearHistory}} (line 127), so with one memory id per user or
session the map and each history grow for the life of the endpoint. Read in the
code, not run.
Proposed: bound the map (size or idle time) and the history length, or use a
pluggable chat memory store.
h4. 2. Any output containing the quoted text "error" is treated as a failure
Every predictor checks {{result.contains("\"error\"")}} (for example
{{TextGenerationPredictor}} line 107, {{ChatPredictor}} line 169,
{{SummarizationPredictor}} line 109) and throws "Python inference failed", so a
generated text that contains the word {{"error"}} in quotes (for example a JSON
snippet) fails. Checked on 4.24.0-SNAPSHOT by calling the predictors'
{{processOutput}} directly.
Proposed: parse the JSON and check for a top-level {{error}} key only.
h4. 3. text-to-speech sets Content-Type application/octet-stream on a POJO body
{{TextToSpeechPredictor}} (line 136) sets {{Content-Type:
application/octet-stream}}, but the body is a DJL
{{ai.djl.modality.audio.Audio}} object, not bytes. Checked on 4.24.0-SNAPSHOT
by calling the predictor directly.
Proposed: set the body to the encoded audio bytes (WAV) to match the header, or
do not set the header.
_Claude Code on behalf of davsclaus_
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
