Spaces:
Running
Running
fix prompts
Browse files
agents/context_compression_agent/context_compression_agent.py
CHANGED
|
@@ -19,4 +19,4 @@ def compress_context(context: List[PlanningStep]):
|
|
| 19 |
}
|
| 20 |
],
|
| 21 |
)
|
| 22 |
-
return completion
|
|
|
|
| 19 |
}
|
| 20 |
],
|
| 21 |
)
|
| 22 |
+
return completion.content
|
agents/context_compression_agent/get_context_compression_prompt.py
CHANGED
|
@@ -10,13 +10,23 @@ def get_context_compression_prompt():
|
|
| 10 |
- Remove any information that is not directly relevant to the task at hand.
|
| 11 |
- You can also remove past planning phases, since the agent has a plan that is saved on disk.
|
| 12 |
- Do NOT add new information or assumptions. Never guess missing data.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
Compressed output rules:
|
| 15 |
-
- Use short, factual bullets or structured key=value notation.
|
| 16 |
- Keep technical names unchanged (tables, fields, tools, IDs, URLs).
|
| 17 |
- Maintain chronological order only if it affects correctness.
|
| 18 |
-
- If information is incomplete,
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
Your goal: **Minimum tokens, maximum retention, zero hallucination.**
|
| 21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
Output:"""
|
|
|
|
| 10 |
- Remove any information that is not directly relevant to the task at hand.
|
| 11 |
- You can also remove past planning phases, since the agent has a plan that is saved on disk.
|
| 12 |
- Do NOT add new information or assumptions. Never guess missing data.
|
| 13 |
+
- Do NOT add any new information or assumptions. Never guess missing data.
|
| 14 |
+
|
| 15 |
+
IMPORTANT TASK:
|
| 16 |
+
- If a big text is repeated several times, define it as a variable in the text and replace occurrences with the variable name to save space.
|
| 17 |
|
| 18 |
Compressed output rules:
|
|
|
|
| 19 |
- Keep technical names unchanged (tables, fields, tools, IDs, URLs).
|
| 20 |
- Maintain chronological order only if it affects correctness.
|
| 21 |
+
- If information is incomplete, do not try to guess it.
|
| 22 |
+
- Keep the same structure as the input, with the same field names. If you receive a list with objects, return a list with objects. KEEP THE EXACT STRUCTURE.
|
| 23 |
+
- Your output will replace the old context, so try to keep the same format and remove things only if it's necessary.
|
| 24 |
+
- Think of your output as the new context that will be passed to the agent.
|
| 25 |
|
| 26 |
Your goal: **Minimum tokens, maximum retention, zero hallucination.**
|
| 27 |
|
| 28 |
+
Output MUST be valid JSON only. Do not add comments, explanations or markdown.
|
| 29 |
+
|
| 30 |
+
YOU MUST PRESERVE ALL KEYS FROM THE INPUT, DO NOT REMOVE KEYS, as this will break the context processing.
|
| 31 |
+
|
| 32 |
Output:"""
|
agents/supabase_agent/get_supabase_agent_prompt.py
CHANGED
|
@@ -72,7 +72,7 @@ def get_supabase_agent_prompt(managed_agent_prompt: dict):
|
|
| 72 |
YOU MUST USE IT instead of writing SQL.
|
| 73 |
|
| 74 |
|
| 75 |
-
|
| 76 |
|
| 77 |
{supabase_agent_return_schema}
|
| 78 |
|
|
|
|
| 72 |
YOU MUST USE IT instead of writing SQL.
|
| 73 |
|
| 74 |
|
| 75 |
+
ALWAYS return the data in a JSON object. The JSON object must match the following schema:
|
| 76 |
|
| 77 |
{supabase_agent_return_schema}
|
| 78 |
|