python-fastllm 0.0.12__tar.gz → 0.0.15__tar.gz

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fastllm
-Version: 0.0.12
+Version: 0.0.15
 Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
 License: Apache-2.0
 Project-URL: Repository, https://github.com/AnswerDotAI/fastllm
@@ -11,6 +11,7 @@ Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 Requires-Dist: fastspec
 Requires-Dist: toolslm
+Requires-Dist: pillow
 
 # fastllm
 

python_fastllm-0.0.15/fastllm/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.0.15"

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/_modidx.py

@@ -19,6 +19,7 @@ d = { 'settings': { 'branch': 'main',
                                    'fastllm.acomplete.mk_client': ('acomplete.html#mk_client', 'fastllm/acomplete.py')},
             'fastllm.anthropic': { 'fastllm.anthropic._ant_cc': ('anthropic.html#_ant_cc', 'fastllm/anthropic.py'),
                                    'fastllm.anthropic._ant_part_type': ('anthropic.html#_ant_part_type', 'fastllm/anthropic.py'),
+                                   'fastllm.anthropic._sanid': ('anthropic.html#_sanid', 'fastllm/anthropic.py'),
                                    'fastllm.anthropic.acollect_stream': ('anthropic.html#acollect_stream', 'fastllm/anthropic.py'),
                                    'fastllm.anthropic.cost': ('anthropic.html#cost', 'fastllm/anthropic.py'),
                                    'fastllm.anthropic.delta_index_fn': ('anthropic.html#delta_index_fn', 'fastllm/anthropic.py'),
@@ -204,7 +205,8 @@ d = { 'settings': { 'branch': 'main',
                                      'fastllm.openai_chat.norm_parts': ('oai_chat.html#norm_parts', 'fastllm/openai_chat.py'),
                                      'fastllm.openai_chat.norm_sse_event': ('oai_chat.html#norm_sse_event', 'fastllm/openai_chat.py'),
                                      'fastllm.openai_chat.norm_tool_calls': ('oai_chat.html#norm_tool_calls', 'fastllm/openai_chat.py')},
-            'fastllm.openai_responses': { 'fastllm.openai_responses.acollect_stream': ( 'oai_responses.html#acollect_stream',
+            'fastllm.openai_responses': { 'fastllm.openai_responses._sanid': ('oai_responses.html#_sanid', 'fastllm/openai_responses.py'),
+                                          'fastllm.openai_responses.acollect_stream': ( 'oai_responses.html#acollect_stream',
                                                                                         'fastllm/openai_responses.py'),
                                           'fastllm.openai_responses.cost': ('oai_responses.html#cost', 'fastllm/openai_responses.py'),
                                           'fastllm.openai_responses.delta_index_fn': ( 'oai_responses.html#delta_index_fn',
@@ -291,5 +293,6 @@ d = { 'settings': { 'branch': 'main',
                                'fastllm.types.part_txt': ('types.html#part_txt', 'fastllm/types.py'),
                                'fastllm.types.payload_kwargs': ('types.html#payload_kwargs', 'fastllm/types.py'),
                                'fastllm.types.register_model_info': ('types.html#register_model_info', 'fastllm/types.py'),
+                               'fastllm.types.resize_b64': ('types.html#resize_b64', 'fastllm/types.py'),
                                'fastllm.types.sys_text': ('types.html#sys_text', 'fastllm/types.py'),
                                'fastllm.types.url_mime': ('types.html#url_mime', 'fastllm/types.py')}}}

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/anthropic.py

@@ -85,6 +85,7 @@ def norm_parts(resp):
 
 # %% ../nbs/04_anthropic.ipynb #a3869e31
 def norm_sse_event(ev, **kwargs):
+    ev = obj2dict(ev)
     typ = ev.get("type")
     text, thinking, tcs, citations = None, None, [], None
     if typ == "content_block_start":
@@ -125,11 +126,14 @@ def _ant_cc(block, p):
     if (cc := (p.data or {}).get('cache_control')): block['cache_control'] = cc
     return block
 
+# %% ../nbs/04_anthropic.ipynb #62e9a042
+def _sanid(id_str): return re.sub(r'[^a-zA-Z0-9_-]', '_', id_str or '')
+
 # %% ../nbs/04_anthropic.ipynb #6ec772cb
 def denorm_tool_use(p:Part):
     "Convert canonical tool_use Part to Anthropic tool_use content block."
     d = p.data or {}
-    block = dict(type='tool_use', id=d.get('id',''), name=d.get('name',''), input=d.get('arguments') or {})
+    block = dict(type='tool_use', id=_sanid(d.get('id','')), name=d.get('name',''), input=d.get('arguments') or {})
     if 'caller' in d: block['caller'] = d['caller']
     return _ant_cc(block, p)
 
@@ -238,8 +242,10 @@ def denorm_user(m:Msg):
     return dict(role='user', content=parts)
 
 # %% ../nbs/04_anthropic.ipynb #edd87272
-def denorm_image(p):
-    if (b64:=data_url(p.text)): return {"type": "image", "source": {"type": "base64", "media_type": b64[0], "data": b64[1]}}
+def denorm_image(p, max_sz=None):
+    if (b64:=data_url(p.text)):
+        data = resize_b64(b64[1], max_sz) if max_sz else b64[1]
+        return {"type": "image", "source": {"type": "base64", "media_type": b64[0], "data": data}}
     return {"type": "image", "source": {"type": "url", "url": p.text}}
 
 # %% ../nbs/04_anthropic.ipynb #fc6bbdfc
@@ -251,12 +257,12 @@ def denorm_file(p):
 def denorm_tool_result(p:Part):
     "Convert canonical tool_result Part to Anthropic tool_result content block."
     d = p.data or {}
-    tid = d.get('id') or d.get('call_id','')
+    tid = _sanid(d.get('id') or d.get('call_id',''))
     if isinstance(p.text, list):
         blocks = []
         for pp in p.text:
             if   pp.type == PartType.text:        blocks.append({"type": "text", "text": pp.text or ""})
-            elif pp.type == PartType.input_image: blocks.append(denorm_image(pp))
+            elif pp.type == PartType.input_image: blocks.append(denorm_image(pp, max_sz=2000))
             elif pp.type == PartType.input_file:  blocks.append(denorm_file(pp))
             else: raise ValueError(f"Anthropic tool_result does not support {pp.type}")
         return _ant_cc(dict(type='tool_result', tool_use_id=tid, content=blocks), p)

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/chat.py

@@ -183,7 +183,7 @@ def fmt2hist(outp:str)->list[Msg]:
     "Transform a formatted output string into fastllm canonical Msgs"
     if token_dtls_tag in outp: outp = re_token.sub('', outp)
     if tool_dtls_tag not in outp:
-        msg = Msg(role='assistant', content=[Part(type=PartType.text, text=outp.strip())])
+        msg = Msg(role='assistant', content=[Part(type=PartType.text, text=outp.strip() or '.')])
         return _split_msg_on_fences(msg)
     hist, asst_parts, tool_parts = [], [], []
     def flush():
@@ -194,7 +194,7 @@ def fmt2hist(outp:str)->list[Msg]:
     for txt,_,tj in split_tools(outp):
         if txt and txt.strip():
             if tool_parts: flush()
-            asst_parts.append(Part(type=PartType.text, text=txt.strip()))
+            asst_parts.append(Part(type=PartType.text, text=txt.strip() or '.'))
         if tj and (tp := _extract_tool_parts(tj)):
             asst_parts.append(tp[0])
             tool_parts.append(tp[1])
@@ -316,7 +316,7 @@ def _has_stop(tres_parts): return any(isinstance(p.text, StopResponse) for p in
 def _trunc_str(s, mx=2000, skip=10, replace="TRUNCATED"):
     "Truncate `s` to `mx` chars max, adding `replace` if truncated"
     if not isinstance(s, str): s = str(s)
-    s = s.rstrip()
+    s = type(s)(s.rstrip())
     if len(s)>2 and s[0]=='𝍁' and s[-1]=='𝍁':
         s = s[1:-1]
         if replace: return s

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/gemini.py

@@ -86,6 +86,7 @@ def norm_parts(resp):
 # %% ../nbs/05_gemini.ipynb #9a5024ee
 def norm_sse_event(ev, **kwargs):
     "Normalize Gemini stream event into Delta."
+    ev = obj2dict(ev)
     cand = nested_idx(ev, 'candidates', 0) or {}
     finish_reason = norm_finish(ev)
     parts = nested_idx(cand, 'content', 'parts') or []

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/openai_chat.py

@@ -55,6 +55,7 @@ def norm_parts(resp):
 def norm_sse_event(ev, **kwargs):
     "Normalize a chat completion stream event."
     # usage always arrives as a single final event with choices: []
+    ev = obj2dict(ev)
     fin = nested_idx(ev, 'choices', 0, 'finish_reason')
     tcs = norm_tool_calls(ev, delta=True)
     if (dlt:=nested_idx(ev, 'choices', 0, 'delta')) is not None:

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/openai_responses.py

@@ -82,6 +82,7 @@ def norm_parts(resp):
 # %% ../nbs/02_oai_responses.ipynb #7cd48aa5
 def norm_sse_event(ev, **kwargs):
     "Normalize OpenAI Responses API stream event into Delta."
+    ev = obj2dict(ev)
     typ = ev.get("type")
     if typ == "response.output_text.delta":            return Delta(text=ev.get("delta"), raw=ev, **kwargs)
     if typ == "response.reasoning_text.delta":         return Delta(thinking=ev.get("delta",""), raw=ev, **kwargs)
@@ -109,10 +110,13 @@ async def acollect_stream(resp, **kwargs):
     res = mk_acollect_stream(norm_and_yield(resp, norm_sse_event), index_fn=delta_index_fn, api_name='openai', **kwargs)
     async for o in res: yield o
 
+# %% ../nbs/02_oai_responses.ipynb #9608e813
+def _sanid(id_str): return id_str[:64] # codex max 64 char limit
+
 # %% ../nbs/02_oai_responses.ipynb #b746c82b
 def denorm_tool_use(p:Part):
     "Convert canonical tool_use Part back to OpenAI Responses function_call item."
-    return dict(type='function_call', call_id=p.data.get('id'), name=p.data.get('name'), arguments=json.dumps(p.data.get('arguments', '{}')))
+    return dict(type='function_call', call_id=_sanid(p.data.get('id')), name=p.data.get('name'), arguments=json.dumps(p.data.get('arguments', '{}')))
 
 # %% ../nbs/02_oai_responses.ipynb #8f42adf7
 def denorm_tool(m:Msg):
@@ -208,7 +212,7 @@ def denorm_file(p):
 # %% ../nbs/02_oai_responses.ipynb #145b1c79
 def denorm_tool_result(m:Part):
     "Convert canonical tool result back to OpenAI Responses function_call_output item."
-    cid = m.data.get('id', '') or m.data.get('call_id')
+    cid = _sanid(m.data.get('id', '') or m.data.get('call_id'))
     if isinstance(m.text, list):
         out = []
         for p in m.text:

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/streaming.py

@@ -137,14 +137,10 @@ async def mk_acollect_stream(it, index_fn, model=None, api_name=None, vendor_nam
                 data = {**acc.extra, 'id':acc.id, 'name':acc.name, 'arguments':args, 'server':acc.server}
                 yield Part(type=PartType.tool_use, data=data)
                 # Server tool results for anthropic are yielded in d.server_tool_result by checking injected dummy `_delta`
-                if acc.server and '_delta' not in tc.arguments: yield Part(type=PartType.tool_result, text="Server tool call executed.", data=data)
+                if acc.server: yield Part(type=PartType.tool_result, text="Server tool call executed.", data=data)
         if d.server_tool_result:
             idx = _fidx(d, 'server_tool_result')
             part_accum.parts[idx] = Part(type=typ, data=d.server_tool_result)
-            srv_tc = next((p for p in reversed(list(part_accum.parts.values())) if isinstance(p, ToolCall) and p.server), None)
-            if srv_tc:
-                data = {**srv_tc.extra, 'id':srv_tc.id, 'name':srv_tc.name, 'arguments':srv_tc.arguments, 'server':True}
-                yield Part(type=PartType.tool_result, text="Server tool call executed.", data=data)
         r = _proc(d, 'refusal')
         if r[0]: yield r[0]
         if d.finish_reason: fin = d.finish_reason

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/fastllm/types.py

@@ -7,15 +7,16 @@ __all__ = ['PartType', 'FinishReason', 'api_registry', 'model_prices_url', 'haik
            'gpt54', 'gpt54m', 'gpt55', 'codex54', 'codex54m', 'codex55', 'codex53spark', 'model_info_registry',
            'modern_llm', 'deepseek_v4_common', 'mimo_v25_common', 'codex_pricing', 'Part', 'Msg', 'ToolCall',
            'display_list', 'Usage', 'Completion', 'APIRegistry', 'mk_completion', 'mk_tool_res_msg', 'fn_schema',
-           'sys_text', 'part_txt', 'data_url', 'url_mime', 'payload_kwargs', 'get_api_key', 'model_prices_meta',
-           'infer_api_name', 'get_model_meta', 'register_model_info', 'get_model_info', 'get_model_pricing',
-           'approx_pricing']
+           'sys_text', 'part_txt', 'data_url', 'url_mime', 'payload_kwargs', 'get_api_key', 'resize_b64',
+           'model_prices_meta', 'infer_api_name', 'get_model_meta', 'register_model_info', 'get_model_info',
+           'get_model_pricing', 'approx_pricing']
 
 # %% ../nbs/00_types.ipynb #b4d047fd
-import httpx
+import httpx, base64, io
 from dataclasses import dataclass, field
 from fastcore.net import urljson
 from fastcore.utils import *
+from PIL import Image as PImg
 
 # %% ../nbs/00_types.ipynb #e568bade
 @dataclass
@@ -160,6 +161,7 @@ api_registry = APIRegistry()
 # %% ../nbs/00_types.ipynb #d58a5f96
 def mk_completion(resp, model, api_name, vendor_name):
     "Normalize an api response into Completion."
+    resp = obj2dict(resp)
     api = api_registry.apis[api_name]
     tcs = api.norm_tool_calls(resp)
     parts = api.norm_parts(resp)
@@ -242,11 +244,26 @@ def get_api_key(api_key, default):
     if not key: raise ValueError(f"Missing API key: set environment variable '{default}' or pass `api_key` parameter")
     return key
 
+# %% ../nbs/00_types.ipynb #25e9cd60
+def resize_b64(b64, max_sz):
+    "Resize a base64 image data to a max long edge, preserving aspect ratio."
+    img = PImg.open(io.BytesIO(base64.b64decode(b64)))
+    if max(img.size) < max_sz: return b64
+    img.thumbnail((max_sz,max_sz), PImg.Resampling.LANCZOS)
+    fmt = (img.format or 'PNG').upper()
+    if fmt=='JPG': fmt = 'JPEG'
+    buf = io.BytesIO()
+    img.save(buf, format=fmt)
+    return base64.b64encode(buf.getvalue()).decode()
+
 # %% ../nbs/00_types.ipynb #852adecd
 model_prices_url = 'https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json'
 
-@flexicache(time_policy(24*60*60))
-def model_prices_meta(): return urljson(model_prices_url)
+def model_prices_meta():
+    "Download model prices to the module dir once, then load from disk."
+    p = Path(__file__).parent/'model_prices.json'
+    if not p.exists(): p.write_text(httpx.get(model_prices_url, follow_redirects=True).text)
+    return loads(p.read_text())
 
 # %% ../nbs/00_types.ipynb #68e488d8
 def infer_api_name(model):
@@ -256,7 +273,6 @@ def infer_api_name(model):
     if any(o in model for o in ('gpt','o3-','o4-')): return 'openai'
 
 # %% ../nbs/00_types.ipynb #2f0720c2
-@flexicache(time_policy(24*60*60))
 def get_model_meta(model, vendor_name=None, tfm=noop):
     "Look up cost metadata for `model` from litellm price map, using `vendor_name` prefix if needed."
     vendor_name = ifnone(vendor_name, infer_api_name(model))
@@ -264,7 +280,7 @@ def get_model_meta(model, vendor_name=None, tfm=noop):
     if model in mp: key = model
     elif vendor_name=='gemini' and model.startswith('models/'): key = f"gemini/{model.removeprefix('models/')}"
     elif vendor_name:                                           key = f"{vendor_name}/{model}"
-    return dict2obj(tfm(mp.get(key), model, vendor_name))
+    return dict2obj(tfm(mp.get(key, {}), model, vendor_name))
 
 # %% ../nbs/00_types.ipynb #60607e23
 haik45 = "claude-haiku-4-5"
@@ -342,9 +358,9 @@ register_model_info('deepseek-v4-pro', vendor_name='deepseek', base='deepseek/de
 
 mimo_v25_common = dict(**modern_llm, supports_web_search=True, max_input_tokens=1048576, max_output_tokens=131072, max_tokens=131072)
 
-register_model_info('mimo-v2.5-pro', vendor_name='mimo', **mimo_v25_common, base='deepseek/deepseek-v4-pro',
+register_model_info('mimo-v2.5-pro', vendor_name='mimo', **mimo_v25_common, base='deepseek-v4-pro', base_vendor_name='deepseek',
     input_cost_per_token=0.435e-6, output_cost_per_token=0.87e-6, cache_read_input_token_cost=0.0036e-6, search_context_cost_per_query=0.005)
-register_model_info('mimo-v2.5', vendor_name='mimo', **mimo_v25_common, base='deepseek/deepseek-v4',
+register_model_info('mimo-v2.5', vendor_name='mimo', **mimo_v25_common, base='deepseek-v4-pro', base_vendor_name='deepseek',
     input_cost_per_token=0.14e-6,  output_cost_per_token=0.28e-6, cache_read_input_token_cost=0.0028e-6, search_context_cost_per_query=0.005,
     supports_vision=True, supports_image_input=True)
 

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/pyproject.toml

@@ -15,7 +15,7 @@ classifiers = [
     "Programming Language :: Python :: 3",
     "Programming Language :: Python :: 3 :: Only",
 ]
-dependencies = ['fastspec', 'toolslm']
+dependencies = ['fastspec', 'toolslm', 'pillow']
 
 [project.urls]
 Repository = "https://github.com/AnswerDotAI/fastllm"

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/python_fastllm.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fastllm
-Version: 0.0.12
+Version: 0.0.15
 Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
 License: Apache-2.0
 Project-URL: Repository, https://github.com/AnswerDotAI/fastllm
@@ -11,6 +11,7 @@ Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 Requires-Dist: fastspec
 Requires-Dist: toolslm
+Requires-Dist: pillow
 
 # fastllm
 

{python_fastllm-0.0.12 → python_fastllm-0.0.15}/python_fastllm.egg-info/requires.txt

@@ -1,2 +1,3 @@
 fastspec
 toolslm
+pillow

python_fastllm-0.0.12/fastllm/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.0.12"