Feature/configurable system prompt master

backend/dataset/admin.py

@@ -1,4 +1,7 @@
-import resource
+try:
+    import resource
+except ImportError:
+    resource = None
 from django.contrib import admin
 from import_export.admin import ImportExportActionModelAdmin
 from .resources import *

backend/functions/views.py

@@ -365,14 +365,18 @@ def chat_log(request):
 @permission_classes([AllowAny])
 @api_view(["POST"])
 def chat_output(request):
+    DEFAULT_SYSTEM_PROMPT = (
+        "We will be rendering your response on a frontend. So, please add spaces or indentation or nextline chars or "
+        "bullet or numberings etc. suitably for code or the text, wherever required."
+    )
     prompt = request.data.get("message")
     history = request.data.get("history", "")
     model = request.data.get("model", "GPT3.5")
+    system_prompt = (request.data.get("system_prompt", "") or "").strip() or DEFAULT_SYSTEM_PROMPT
     return Response(
         {
             "message": get_model_output(
-                "We will be rendering your response on a frontend. so please add spaces or indentation or nextline chars or "
-                "bullet or numberings etc. suitably for code or the text. wherever required.",
+                system_prompt,
                 prompt,
                 history,
                 model,

backend/projects/models.py

@@ -281,6 +281,8 @@ class Project(models.Model):
         ),
     )
 
+
+
     def clear_expired_lock(self):
         self.lock.filter(expires_at__lt=now()).delete()
 

backend/tasks/views.py

@@ -1812,7 +1812,7 @@ def partial_update(self, request, pk=None):
                                 annotation_obj.task,
                                 annotation_obj,
                                 annotation_obj.task.project_id.metadata_json,
-                                task.data["model"]
+                                task.data.get("model", [])
                             )
                             if output_result == -1:
                                 ret_dict = {
@@ -2074,7 +2074,7 @@ def partial_update(self, request, pk=None):
                                 annotation_obj.task,
                                 annotation_obj,
                                 annotation_obj.task.project_id.metadata_json,
-                                task.data["model"]
+                                task.data.get("model", [])
                             )
                             if output_result == -1:
                                 ret_dict = {
@@ -2405,7 +2405,7 @@ def partial_update(self, request, pk=None):
                                 annotation_obj.task,
                                 annotation_obj,
                                 annotation_obj.task.project_id.metadata_json,
-                                task.data["model"]
+                                task.data.get("model", [])
                             )
                             if output_result == -1:
                                 ret_dict = {
@@ -2938,16 +2938,12 @@ def get_llm_output(prompt, task, annotation, project_metadata_json):
     intent = task.data["meta_info_intent"]
     domain = task.data["meta_info_domain"]
     lang_type = task.data["meta_info_language"]
-    ann_result = (
-        json.loads(annotation.result)
-        if isinstance(annotation.result, str)
-        else annotation.result
-    )
-    project_metadata = (
-        json.loads(project_metadata_json)
-        if isinstance(project_metadata_json, str)
-        else project_metadata_json
-    )
+    ann_result = annotation.result
+    if isinstance(ann_result, str):
+        ann_result = json.loads(ann_result) if ann_result.strip() else []
+    project_metadata = project_metadata_json
+    if isinstance(project_metadata, str):
+        project_metadata = json.loads(project_metadata) if project_metadata.strip() else {}
     if isinstance(project_metadata, dict) and project_metadata.get("blank_response") == True:
         return ""
     if prompt in [None, "Null", 0, "None", "", " "]:
@@ -2980,11 +2976,21 @@ def get_llm_output(prompt, task, annotation, project_metadata_json):
         dup_check = duplicate_check(ann_result, prompt)
 
     # GET MODEL OUTPUT
-    history = ann_result
+    DEFAULT_SYSTEM_PROMPT = (
+        "We will be rendering your response on a frontend. So, please add spaces or indentation or nextline chars or "
+        "bullet or numberings etc. suitably for code or the text, wherever required."
+    )
+    sys_prompt_data = project_metadata.get("system_prompt", {}) if isinstance(project_metadata, dict) else {}
     model = task.data["model"]
+
+    if isinstance(sys_prompt_data, dict):
+        system_prompt = sys_prompt_data.get(model) or sys_prompt_data.get("default") or DEFAULT_SYSTEM_PROMPT
+    else:
+        system_prompt = sys_prompt_data.strip() if sys_prompt_data.strip() else DEFAULT_SYSTEM_PROMPT
+
+    history = ann_result
     model_output = get_model_output(
-        "We will be rendering your response on a frontend. so please add spaces or indentation or nextline chars or "
-        "bullet or numberings etc. suitably for code or the text. wherever required.",
+        system_prompt,
         prompt,
         history,
         model,
@@ -2995,20 +3001,23 @@ def get_llm_output(prompt, task, annotation, project_metadata_json):
     return res
 
 def get_all_llm_output(prompt, task, annotation, project_metadata_json, models_to_run):
+    project_metadata = project_metadata_json
+    if isinstance(project_metadata, str):
+        project_metadata = json.loads(project_metadata) if project_metadata.strip() else {}
+
+    if not models_to_run:
+        models_to_run = project_metadata.get("models_set", []) if isinstance(project_metadata, dict) else []
+
+    if not models_to_run:
+        return Response({"message": "No models are configured for this task. Please configure models for this project."}, status=status.HTTP_400_BAD_REQUEST)
+
     # CHECKS
-    intent = task.data["meta_info_intent"]
-    domain = task.data["meta_info_domain"]
-    lang_type = task.data["meta_info_language"]
-    ann_result = (
-        json.loads(annotation.result)
-        if isinstance(annotation.result, str)
-        else annotation.result
-    )
-    project_metadata = (
-        json.loads(project_metadata_json)
-        if isinstance(project_metadata_json, str)
-        else project_metadata_json
-    )
+    intent = task.data.get("meta_info_intent")
+    domain = task.data.get("meta_info_domain")
+    lang_type = task.data.get("meta_info_language")
+    ann_result = annotation.result
+    if isinstance(ann_result, str):
+        ann_result = json.loads(ann_result) if ann_result.strip() else []
     if prompt in [None, "Null", 0, "None", "", " "]:
         return -1
     intentDomain_test, lang_test, duplicate_test = False, False, False
@@ -3039,15 +3048,24 @@ def get_all_llm_output(prompt, task, annotation, project_metadata_json, models_t
         dup_check = duplicate_check(ann_result, prompt)
 
     # GET MODEL OUTPUT
+    DEFAULT_SYSTEM_PROMPT = (
+        "We will be rendering your response on a frontend. So, please add spaces or indentation or nextline chars or "
+        "bullet or numberings etc. suitably for code or the text, wherever required."
+    )
+    sys_prompt_data = project_metadata.get("system_prompt", {}) if isinstance(project_metadata, dict) else {}
+
+    if not isinstance(sys_prompt_data, dict):
+        sys_prompt_data = {"default": sys_prompt_data.strip() if sys_prompt_data.strip() else DEFAULT_SYSTEM_PROMPT}
+
     history = ann_result[0]
 
 
     model_output = get_all_model_output(
-        "We will be rendering your response on a frontend. so please add spaces or indentation or nextline chars or "
-        "bullet or numberings etc. suitably for code or the text. wherever required.",
+        sys_prompt_data,
         prompt,
         history,
-        models_to_run
+        models_to_run,
+        DEFAULT_SYSTEM_PROMPT
     )
 
     return model_output

backend/utils/llm_interactions.py

@@ -161,9 +161,16 @@ def get_llama2_output(system_prompt, conv_history, user_prompt):
         "max_new_tokens": 500,
         "top_p": 1,
     }
-    s = requests.Session()
-    result = s.post(url, headers={"Authorization": f"Bearer {token}"}, json=body)
-    return result.json()["choices"][0]["message"]["content"].strip()
+    try:
+        s = requests.Session()
+        result = s.post(url, headers={"Authorization": f"Bearer {token}"}, json=body)
+        result.raise_for_status()
+        return result.json()["choices"][0]["message"]["content"].strip()
+    except Exception as e:
+        err_msg = str(e)
+        message = f"An error occurred while interacting with Llama2 API: {err_msg}"
+        st = status.HTTP_500_INTERNAL_SERVER_ERROR
+        return Response({"message": message}, status=st)
 
 def get_sarvam_m_output(system_prompt, conv_history, user_prompt):
     api_base = os.getenv("SARVAM_M_API_BASE")
@@ -256,10 +263,11 @@ def get_model_output(system_prompt, user_prompt, history, model=GPT4OMini):
         out = get_deepinfra_output(system_prompt, user_prompt, history, model)
     return out
 
-def get_all_model_output(system_prompt, user_prompt, history, models_to_run):
+def get_all_model_output(system_prompt_data, user_prompt, history, models_to_run, default_system_prompt=""):
     results = {}
 
     for model in models_to_run:
+        system_prompt = system_prompt_data.get(model) or system_prompt_data.get("default") or default_system_prompt if isinstance(system_prompt_data, dict) else system_prompt_data
         # print("history:", history)
         # model_history = next(
         #     (entry["interaction_json"] for entry in history if entry.get("model_name") == model),
@@ -284,4 +292,7 @@ def get_all_model_output(system_prompt, user_prompt, history, models_to_run):
         else:
             results[model] = get_deepinfra_output(system_prompt, user_prompt, model_history, model)
 
+        if isinstance(results[model], Response):
+            return results[model]
+
     return results