python-fastllm 0.0.60__py3-none-any.whl → 0.0.62__py3-none-any.whl

fastllm/__init__.py

@@ -2,4 +2,4 @@
 
 - `fastllm.responses`: Provider-independent Responses resources and streaming over FastLLM"""
 
-__version__ = "0.0.60"
+__version__ = "0.0.62"

fastllm/acomplete.py

@@ -33,6 +33,7 @@ vendor_mapping = {
     "codex":        ('openai', 'https://chatgpt.com/backend-api/codex', 'CODEX_AUTH_TOKEN', _codex_json),
     "moonshot":     ('openai_chat', "https://api.moonshot.ai/v1", "MOONSHOT_API_KEY"),
     "deepseek":     ('openai_chat', "https://api.deepseek.com/v1", "DEEPSEEK_API_KEY"),
+    "inception":    ('openai_chat', "https://api.inceptionlabs.ai/v1", "INCEPTION_API_KEY"),
     "mimo":         ('anthropic', "https://api.xiaomimimo.com/anthropic", "MIMO_API_KEY"),
     "openrouter":   ('openai_chat', "https://openrouter.ai/api/v1", "OPENROUTER_API_KEY"),
     "together":     ('openai_chat', "https://api.together.xyz/v1", "TOGETHER_API_KEY"),
@@ -167,20 +168,23 @@ async def acomplete(msgs, model, api_name=None, vendor_name=None, api_key=None,
     if own:
         vendor_name = ifnone(vendor_name, api_name)  # an own-transport api: no HTTP client to build
         if oauth_token: kwargs['oauth_token'] = oauth_token  # its `mk_payload` applies the user's token itself
-    else: cli, api_name, vendor_name = mk_client(model=model, vendor_name=vendor_name, api_name=api_name, api_key=api_key, oauth_token=oauth_token, base_url=base_url, xtra_hdrs=xtra_hdrs)
+    else: cli, api_name, vendor_name = mk_client(model=model, vendor_name=vendor_name, api_name=api_name, api_key=api_key, oauth_token=oauth_token, base_url=base_url)
     api = api_registry[api_name]
     if previous_response_id is not None:
         if not getattr(api, 'supports_previous_response_id', False):
             raise ValueError('previous_response_id is not supported by this transport')
         kwargs['previous_response_id'] = previous_response_id
     payload = api.mk_payload(msgs, model, stream=stream, **kwargs)
-    xtra_body = ifnone(xtra_body, {})
     if fix := getattr(api, 'fix_payload', None): fix(payload, model, vendor_name)
+    body = payload | ifnone(xtra_body, {})
     if not own:
+        headers = xtra_hdrs or {}
+        if vendor_name == 'codex' and (key := body.get('prompt_cache_key')):
+            if not any(k.lower() == 'session-id' for k in headers): headers = {**headers, 'session-id':key}
         ep = endpoint or api.endpoint
         path,params = ep(model, stream) if callable(ep) else (ep, None)
-    if defaults.debug_mode: _debug_print(model, api_name, vendor_name, payload | xtra_body, None if own else cli._url(path))
-    async def _call(): return await provider_req(cli, path, payload | xtra_body, params=params, stream=stream)
+    if defaults.debug_mode: _debug_print(model, api_name, vendor_name, body, None if own else cli._url(path))
+    async def _call(): return await provider_req(cli, path, body, params=params, stream=stream, headers=headers)
     async def _mk_gen():
         resp = payload if own else await _call()  # an own-transport api's `acollect_stream` makes the request itself, from the payload
         async for o in api.acollect_stream(resp, model=model, vendor_name=vendor_name, stop_callables=stop_callables): yield o

fastllm/chat.py

@@ -9,6 +9,7 @@ __all__ = ['effort', 'contents', 'stop_reason', 'cite_footnote', 'postproc', 'Re
            'StopSequencesCallback', 'StreamAccum', 'adisplay_stream']
 
 # %% ../nbs/07_chat.ipynb #d5a3bc1f
+import secrets
 from typing import Optional,Callable
 from fastcore.funccall import mk_ns, call_func, call_func_async, get_schema
 from fastcore.utils import *
@@ -168,6 +169,7 @@ class AsyncChat:
         cache=False,              # Anthropic prompt caching
         cache_idxs:list=[-1],     # Anthropic cache breakpoint idxs, use `0` for sys prompt if provided
         ttl=None,                 # Anthropic prompt caching ttl
+        prompt_cache_key=None,     # Stable conversation key for OpenAI Responses; generated once if omitted
         api_name=None,            # API to use, one of ApiName: openai (responses), openai_chat, anthropic, gemini
         vendor_name=None,         # Vendor name, one of vendor_mapping which resolves api_base/api_key automatically
         api_key=None,             # API key when model can't be resolved or vendor_name is not known or codex
@@ -195,6 +197,7 @@ class AsyncChat:
         self.response_id,self._response_hist_idx = None,0
         self._turn_start = 0    # index into `hist` where the current turn began, for `full`
         self.last_req_use = None  # usage of the latest request only; `use` accumulates across a turn's tool-call steps
+        if prompt_cache_key is None: prompt_cache_key = secrets.token_hex(16)
         store_attr(but='cbs')
         self.cbs = L()
         if default_cbs: self.add_cbs(defaults.chat_callbacks)
@@ -256,6 +259,7 @@ def _prep_call(self:AsyncChat, search, max_tokens, kwargs, stream=False, think=N
     if self.base_url:      kwargs['base_url'] = self.base_url
     if self.endpoint:      kwargs['endpoint'] = self.endpoint
     if self.extra_headers: kwargs['xtra_hdrs'] = self.extra_headers
+    kwargs.setdefault('prompt_cache_key', self.prompt_cache_key)
     kwargs.update(_think_kw(self.model, think, self.vendor_name))
     return max_tokens