Auto-update: Thu Aug 8 05:13:21 PDT 2024

sijapi/__init__.py

@@ -90,7 +90,7 @@ SUMMARY_CHUNK_OVERLAP = int(os.getenv("SUMMARY_CHUNK_OVERLAP", 256))  # measured
 SUMMARY_TPW = float(os.getenv("SUMMARY_TPW", 1.3))  # measured in tokens
 SUMMARY_LENGTH_RATIO = int(os.getenv("SUMMARY_LENGTH_RATIO", 4))  # measured as original to length ratio
 SUMMARY_MIN_LENGTH = int(os.getenv("SUMMARY_MIN_LENGTH", 150))  # measured in tokens
-SUMMARY_MODEL = os.getenv("SUMMARY_MODEL", "dolphin-llama3:8b-256k")
+SUMMARY_MODEL = os.getenv("SUMMARY_MODEL", "command-r:latest")
 SUMMARY_TOKEN_LIMIT = int(os.getenv("SUMMARY_TOKEN_LIMIT", 16384))
 SUMMARY_INSTRUCT = os.getenv('SUMMARY_INSTRUCT', "You are an AI assistant that provides accurate summaries of text -- nothing more and nothing less. You must not include ANY extraneous text other than the sumary. Do not include comments apart from the summary, do not preface the summary, and do not provide any form of postscript. Do not add paragraph breaks. Do not add any kind of formatting. Your response should begin with, consist of, and end with an accurate plaintext summary.")
 SUMMARY_INSTRUCT_TTS = os.getenv('SUMMARY_INSTRUCT_TTS', "You are an AI assistant that provides email summaries for Sanjay. Your response will undergo Text-To-Speech conversion and added to Sanjay's private podcast. Providing adequate context (Sanjay did not send this question to you, he will only hear your response) but aiming for conciseness and precision, and bearing in mind the Text-To-Speech conversion (avoiding acronyms and formalities), summarize the following email.")

sijapi/config/.env-example

@@ -326,7 +326,7 @@ SYSTEM_MSG=You are a helpful AI assistant.
 DEFAULT_LLM=dolphin-mistral
 DEFAULT_VISION=llava-llama3
 OPENAI_API_KEY=¿SECRET?                # <--- not presently implemented for anything
-SUMMARY_MODEL='dolphin-llama3:8b-256k'
+SUMMARY_MODEL='command-r:latest'
 SUMMARY_CHUNK_SIZE=16384
 SUMMARY_CHUNK_OVERLAP=100
 SUMMARY_TPW=1.3

sijapi/config/img.yaml-example

@@ -7,7 +7,7 @@ scenes:
     API_NegativePrompt: "3d, bad art, illustrated, deformed, blurry, duplicate, video game, render, anime, cartoon, fake, tiling, out of frame, bad art, bad anatomy, 3d render, nsfw, worst quality, low quality, text, watermark, Thomas Kinkade, sentimental, kitsch, kitschy, twee, commercial, holiday card, comic, cartoon"
     llm_sys_msg: "You are a helpful AI who assists in generating prompts that will be used to generate highly realistic images. Distill the described image or scene to its visual essence, in the form of the most evocative and visually descriptive words, phrases, and sentence fragments. Emphasize tone, lighting, composition, and any interesting juxtapositions between foreground and background, or center of frame and outer frame areas. Strive for nuance and originality. Avoid cliches, common tropes, and sentimentality. Avoid vague or abstract concepts. Avoid any words or descriptions based on emotion or any senses besides vision. Strive to show rather than tell. Space is limited, so be efficient with your words."
     llm_pre_prompt: "Using the most visually descriptive sentence fragments, phrases, and words, distill the scene description to its essence, staying true to what it describes: "
-    prompt_model: "dolphin-llama3:70b"
+    prompt_model: "command-r:latest"
     workflows:
       - workflow: default.json
         size: 1024x768
@@ -19,7 +19,7 @@ scenes:
     API_NegativePrompt: "3d, bad art, illustrated, deformed, blurry, duplicate, video game, render, anime, cartoon, fake, tiling, out of frame, bad art, bad anatomy, 3d render, nsfw, worst quality, low quality, text, watermark, Thomas Kinkade, sentimental, kitsch, kitschy, twee, commercial, holiday card, comic, cartoon"
     llm_sys_msg: "You are a helpful AI who assists in generating prompts that will be used to generate highly realistic images. Distill the described image or scene to its visual essence, in the form of the most evocative and visually descriptive words, phrases, and sentence fragments. Emphasize tone, lighting, composition, and any interesting juxtapositions between foreground and background, or center of frame and outer frame areas. Strive for nuance and originality. Avoid cliches, common tropes, and sentimentality. Avoid vague or abstract concepts. Avoid any words or descriptions based on emotion or any senses besides vision. Strive to show rather than tell. Space is limited, so be efficient with your words."
     llm_pre_prompt: "Using the most visually descriptive sentence fragments, phrases, and words, distill the scene description to its essence, staying true to what it describes: "
-    prompt_model: "dolphin-llama3:70b"
+    prompt_model: "command-r:latest"
     workflows:
       - workflow: wallpaper.json
         size: 1024x640
@@ -33,7 +33,7 @@ scenes:
     API_NegativePrompt: "canvas frame, 3d, bad art, illustrated, deformed, blurry, duplicate, bad anatomy, worst quality, low quality, watermark, FastNegativeV2, easynegative, epiCNegative, easynegative, verybadimagenegative_v1.3, nsfw, nude"
     llm_sys_msg: "You are a helpful AI who assists in refining prompts that will be used to generate highly realistic portrait photos. Upon receiving a prompt, you refine it by simplifying and distilling it to its essence, retaining the most visually evocative and distinct elements from what was provided, focusing in particular on the pictured individual's eyes, pose, and other distinctive features. You may infer some visual details that were not provided in the prompt, so long as they are consistent with the rest of the prompt. Always use the most visually descriptive terms possible, and avoid any vague or abstract concepts. Do not include any words or descriptions based on other senses or emotions. Strive to show rather than tell. Space is limited, so be efficient with your words. Remember that the final product will be a still image, and action verbs are not as helpful as simple descriptions of position, appearance, background, etc."
     llm_pre_prompt: "Using the most visually descriptive sentence fragments, phrases, and words, distill this portrait photo to its essence: "
-    prompt_model: "dolphin-llama3:70b"
+    prompt_model: "command-r:latest"
     workflows:
       - workflow: selfie.json
         size: 768x1024

sijapi/config/llm.yaml-example

@@ -6,7 +6,7 @@ chat:
 vision:
   model: llava-llama3
 summary:
-  model: dolphin-llama3:8b-256k
+  model: 'command-r:latest'
   chunk-size: 16384
   chunk-overlap: 256
   length-ratio: 4

sijapi/utilities.py

@@ -52,31 +52,6 @@ def warn(text: str): logger.warning(text)
 def err(text: str): logger.error(text)
 def crit(text: str): logger.critical(text)
 
-api_key_header = APIKeyHeader(name="Authorization", auto_error=False)
-
-def validate_api_key(request: Request, api_key: str = Depends(api_key_header)):
-    if request.url.path in API.PUBLIC:
-        return
-
-    client_ip = ipaddress.ip_address(request.client.host)
-    trusted_subnets = [ipaddress.ip_network(subnet) for subnet in API.TRUSTED_SUBNETS]
-    if any(client_ip in subnet for subnet in trusted_subnets):
-        return
-
-    # Check header-based API key
-    if api_key:
-        if api_key.lower().startswith("bearer "):
-            api_key = api_key.lower().split("bearer ")[-1]
-        if api_key in API.KEYS:
-            return
-
-    # Check query-based API key
-    api_key_query = request.query_params.get("api_key")
-    if api_key_query in API.KEYS:
-        return
-
-    raise HTTPException(status_code=401, detail="Invalid or missing API key")
-
 
 def assemble_archive_path(filename: str, extension: str = None, date_time: dt_datetime = None, subdir: str = None) -> Tuple[Path, Path]:
     date_time = date_time or dt_datetime.now()