Skip to content

CHANGES_SUMMARY.md

Latest commit

 

History

History
96 lines (70 loc) · 3.01 KB

CHANGES_SUMMARY.md

File metadata and controls

96 lines (70 loc) · 3.01 KB

Movie Selection Fix - Changes Summary

Issues Found from Logs

From hal_voice_assistant.log at 03:16:48:

  • User asked: "Play movie about barbarian"
  • System selected: Millennium.mp4 (WRONG)
  • Should have selected: Red Sonja.mp4 (CORRECT)

Root Causes

  1. Pseudo-function interception was broken (FIXED)

    • Code only checked for "select_movie"
    • Agent called "media-server.select_movie" (with server prefix)
    • Interception failed → tried to call as real MCP tool → "Unknown tool"

  2. Selection prompt was too complex (FIXED)

    • Original had numbered lists, multiple IMPORTANT instructions
    • Too much structure confused the small model
    • User tested exaone3.5:2.4b with simple prompt → worked correctly!

Changes Made

1. Fixed Pseudo-function Interception

File: hal_voice_assistant.py:387

Before:

if function_name == "select_movie":

After:

if function_name == "select_movie" or function_name.endswith(".select_movie"):

2. Simplified Selection Prompt

File: hal_voice_assistant.py:410-415

Before:

The user requested: "Play movie about barbarian."

Available movies:
1. Airplane!.mp4
2. Millennium.mp4
3. Nobody.mp4
4. Red Sonja.mp4
5. Superman.mp4

Based ONLY on the movie titles above, which movie best matches the user's request?

IMPORTANT: Choose from the list above. Do NOT make up movies. Consider what each title suggests the movie is about.

Respond with ONLY the exact filename from the list (e.g., "Red Sonja.mp4"), nothing else.

After:

Of the movies "Airplane!.mp4", "Millennium.mp4", "Nobody.mp4", "Red Sonja.mp4", "Superman.mp4", which one best matches: Play movie about barbarian.

Respond with only the filename from the list above, nothing else.

Why: User tested exaone3.5:2.4b and found simpler prompts work better. The model correctly identified Red Sonja when given: "of the movies "superman","airplane","red sonja" which one is about a barbarian"

3. Enhanced Logging

File: hal_voice_assistant.py:421-435

Now logs:

  • Full selection prompt sent to LLM
  • Model name used for selection
  • Complete LLM response (not just extracted filename)

4. Fixed Example Movie Names in System Prompt

File: hal_voice_assistant.py:322-324

Before: Used "Superman.mp4" as example (could be a real movie) After: Uses "nonexistantmovie.mp4" (canary value)

If "nonexistantmovie.mp4" is ever selected, we know the agent is using examples from the prompt instead of real data.

Testing

Ready to test! The new prompt format matches the user's successful test case.

Expected behavior:

  1. User: "play movie about barbarian"
  2. Intent matcher detects "about" → triggers agent mode
  3. Agent calls list_movies → gets movie list
  4. Agent calls select_movie (or media-server.select_movie)
  5. Code intercepts (now works with server prefix!)
  6. Simplified prompt sent to LLM
  7. Should select: Red Sonja.mp4
  8. Agent calls play_movie with correct filename
  9. Logs show full selection prompt and reasoning