python-fastllm 0.0.58__tar.gz → 0.0.60__tar.gz

python_fastllm-0.0.60/PKG-INFO

@@ -0,0 +1,231 @@
+Metadata-Version: 2.4
+Name: python-fastllm
+Version: 0.0.60
+Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
+License: Apache-2.0
+Project-URL: Repository, https://github.com/AnswerDotAI/fastllm
+Keywords: nbdev
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+Requires-Dist: fastcore>=2.1.18
+Requires-Dist: aidialog>=0.0.27
+Requires-Dist: fasttransport>=0.0.2
+Requires-Dist: pillow
+
+# fastllm
+
+
+<!-- WARNING: THIS FILE WAS AUTOGENERATED! DO NOT EDIT! -->
+
+`fastllm` provides a common async interface for Anthropic, OpenAI Responses and Chat, Gemini, and OpenAI-compatible providers such as Kimi. Define messages and tools in a shared format and call `acomplete`. Select the model and, where needed, its provider without rewriting the conversation or tool loop. Supported features vary by provider.
+
+`AsyncResponses` provides a streaming Responses API interface over the same providers. It translates input items and events and keeps a canonical `ResponseState` between turns. For transports without provider continuation ids, it replays that state automatically.
+
+## Install
+
+Run `pip install python-fastllm` or clone from github and install locally.
+
+## Setup
+
+``` python
+from aidialog.msg_parts import Msg, Part, Text, Thinking, ToolUse, InputImage, mk_tool_res_msg, Completion
+from fastllm.acomplete import acomplete
+from fastllm.types import sample_img_url, sample_doms
+import asyncio, json
+```
+
+``` python
+def user(text): return Msg(role='user', content=[Text(text)])
+
+async def stream(msgs, model, max_think=0, **kw):
+    "Print streamed parts and return the final Completion; limit printed thinking deltas with max_think."
+    cnt = 0
+    async for o in await acomplete(msgs, model, stream=True, **kw):
+        if not isinstance(o, Part): continue
+        if isinstance(o, Thinking):
+            cnt += 1
+            if cnt > max_think: continue
+        print(o.formatted, end='', flush=True)
+    print()
+    return o
+```
+
+``` python
+mtok = 1024
+```
+
+## Chat across providers
+
+Use the same `acomplete` interface with Claude, GPT, Gemini, or Kimi. Set `model` and any provider-specific routing arguments:
+
+``` python
+models = [
+    ('claude-sonnet-5', {}),
+    ('gpt-4o-mini', {}),
+    ('models/gemini-3-flash-preview', {}),
+    ('accounts/fireworks/models/kimi-k3', dict(vendor_name='fireworks_ai'))
+]
+for name, kw in models:
+    r = await acomplete([user("Translate 'hello' into French. Return only the translation.")],
+                       model=name, max_tokens=mtok, **kw)
+    text = ''.join(p.text for p in r.message.content if isinstance(p, Text))
+    print(f"{name}: {text.strip()}")
+```
+
+    claude-sonnet-5: Bonjour
+    gpt-4o-mini: Bonjour
+    models/gemini-3-flash-preview: Bonjour
+    accounts/fireworks/models/kimi-k3: Bonjour
+
+## System prompts
+
+Pass `system` to supply a system prompt. `fastllm` translates it to the provider’s field: Anthropic `system`, OpenAI Responses `instructions`, or Gemini `system_instruction`.
+
+``` python
+sys = "You are a pirate chef. Always respond in pirate speak and mention food. Use one short sentence."
+
+print("Claude: ", end='')
+r = await stream([user("What should I do today?")], model='claude-sonnet-5', system=sys, max_tokens=mtok)
+
+print("Gemini: ", end='')
+r = await stream([user("What should I do today?")], model='models/gemini-3-flash-preview', system=sys, max_tokens=mtok)
+```
+
+    Claude: Arrr, chart a course fer the galley and cook up a hearty stew, matey!
+    Gemini: Sharpen yer cutlass and feast on spicy shark stew, ye scurvy dog!
+
+## Tool calling
+
+Define tools in the shared schema. `fastllm` translates it to the provider’s tool format. This example handles a tool request, adds the result to the history, and continues the conversation:
+
+``` python
+tools = [{"type": "function", "function": {
+    "name": "get_weather",
+    "description": "Get current weather for a city",
+    "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
+}}]
+
+msgs = [user("What's the weather in Paris?")]
+r1 = await stream(msgs, model='claude-sonnet-5', tools=tools, max_tokens=mtok)
+print("Tool calls:", r1.tool_calls)
+```
+
+    - ⏳ `get_weather(city="Paris")` ⏳
+
+    Tool calls: [ToolUse(raw={'caller': {'type': 'direct'}}, cache_control=None, id='toolu_01XacBpJHnsjtPWJWr55nA2m', name='get_weather', arguments={'city': 'Paris'}, server=False, text=None)]
+
+``` python
+msgs += [r1.message, mk_tool_res_msg(r1.tool_calls, ['22°C, sunny with light clouds']),
+         user("Should I bring a jacket? Answer in one short sentence.")]
+r2 = await stream(msgs, model='claude-sonnet-5', tools=tools, max_tokens=mtok)
+```
+
+    No, a jacket isn't necessary — it's a mild, sunny 22°C day.
+
+## Tool Choice
+
+Control whether the model must use tools, can’t use tools, or decides on its own:
+
+``` python
+r = await acomplete([user("Hello there!")], model='claude-sonnet-5',
+                    tools=tools, tool_choice='required', max_tokens=mtok)
+print("Required:", [tc.name for tc in r.tool_calls])
+
+r = await acomplete([user("What's the weather?")], model='claude-sonnet-5',
+                    tools=tools, tool_choice='none', max_tokens=mtok)
+print("None:", r.tool_calls)
+```
+
+    Required: ['get_weather']
+    None: []
+
+## Thinking and extended reasoning
+
+Set `reasoning_effort` to `low`, `medium`, or `high` for supported models. The adapters translate these values to provider-specific settings. Reasoning output uses `Thinking` parts; this example prints a short excerpt alongside the answer.
+
+``` python
+print("Claude: ", end='')
+r = await stream([user("What is 127 × 849? Return just the number.")], model='claude-sonnet-4-6',
+                 reasoning_effort='low', max_tokens=8192)
+for p in r.message.content:
+    if isinstance(p, Thinking): print(f"Thinking excerpt: {p.text[:150]}...")
+
+print("Kimi: ", end='')
+r = await stream([user("What is 127 × 849? Return just the number.")],
+                 model='accounts/fireworks/models/kimi-k3', vendor_name='fireworks_ai',
+                 reasoning_effort='low', max_tokens=8192)
+for p in r.message.content:
+    if isinstance(p, Thinking): print(f"Thinking excerpt: {p.text[:150]}...")
+```
+
+    Claude: 107823
+    Thinking excerpt: 127 × 849 = 127 × 800 + 127 × 49 = 101600 + 6223 = 107823...
+    Kimi: 107823
+    Thinking excerpt: 127 × 849 = 107823. Let me verify: 127*800=101600; 127*49=6223; total 107823. Return just number....
+
+## Web Search (Server Tools)
+
+OpenAI’s Responses API supports server-side web search. Server tool calls are normalized alongside regular tool calls:
+
+``` python
+ws_tools = [{"type": "web_search_preview"}]
+r = await acomplete([user("What is the latest Python release? Reply with the version number only.")],
+                    model='gpt-4o-mini', tools=ws_tools, max_tokens=512)
+print("Server tools used:", [tc.name for tc in r.tool_calls if tc.server])
+```
+
+    Server tools used: ['web_search']
+
+## Caching with Anthropic
+
+Set a part’s `cache_control` to request prompt caching. Later requests can reuse eligible cached content. The example prints usage for two calls with the same system prompt:
+
+``` python
+long_ctx = "You are an expert on the solar system. " * 200
+system = Text(long_ctx, cache_control={'type': 'ephemeral'})
+
+r1 = await acomplete([user("What is Jupiter's mass? Answer in one sentence.")],
+                     model='claude-sonnet-5', system=system, max_tokens=mtok)
+print("First call: cache creation tokens", r1.usage.cache_creation_tokens)
+
+r2 = await acomplete([user("What is Saturn's mass? Answer in one sentence.")],
+                     model='claude-sonnet-5', system=system, max_tokens=mtok)
+print("Second call: cache read tokens", r2.usage.cached_tokens)
+```
+
+    First call: cache creation tokens 2204
+    Second call: cache read tokens 2204
+
+## Media inputs
+
+Use `InputImage` with a model that supports images. The example sends the same image message to three providers:
+
+``` python
+img_msg = Msg(role='user', content=[
+    InputImage(sample_img_url),
+    Text("List three visible objects, using nouns only.")
+])
+
+for name, kw in [('claude-sonnet-5', {}), ('gpt-4o-mini', {}), ('models/gemini-3-flash-preview', {})]:
+    print(f"{name}: ", end='')
+    r = await stream([img_msg], model=name, max_tokens=mtok, **kw)
+```
+
+    claude-sonnet-5: Mountains, lake, trees
+    gpt-4o-mini: 1. Mountains  
+    2. Trees  
+    3. Lake
+    models/gemini-3-flash-preview: planks, lake, mountains
+
+The media adapters support the following part types:
+
+| Media part   | Anthropic | OpenAI Responses | OpenAI Chat | Gemini |
+|--------------|-----------|------------------|-------------|--------|
+| `InputImage` | Yes       | Yes              | Yes         | Yes    |
+| `InputAudio` | No        | No               | Yes         | Yes    |
+| `InputVideo` | No        | No               | No          | Yes    |
+| `InputFile`  | Yes       | Yes              | Yes         | Yes    |
+
+Media parts accept a URL or base64 data URL as their `text`. An unsupported combination raises `ValueError`.

python_fastllm-0.0.60/README.md

@@ -0,0 +1,215 @@
+# fastllm
+
+
+<!-- WARNING: THIS FILE WAS AUTOGENERATED! DO NOT EDIT! -->
+
+`fastllm` provides a common async interface for Anthropic, OpenAI Responses and Chat, Gemini, and OpenAI-compatible providers such as Kimi. Define messages and tools in a shared format and call `acomplete`. Select the model and, where needed, its provider without rewriting the conversation or tool loop. Supported features vary by provider.
+
+`AsyncResponses` provides a streaming Responses API interface over the same providers. It translates input items and events and keeps a canonical `ResponseState` between turns. For transports without provider continuation ids, it replays that state automatically.
+
+## Install
+
+Run `pip install python-fastllm` or clone from github and install locally.
+
+## Setup
+
+``` python
+from aidialog.msg_parts import Msg, Part, Text, Thinking, ToolUse, InputImage, mk_tool_res_msg, Completion
+from fastllm.acomplete import acomplete
+from fastllm.types import sample_img_url, sample_doms
+import asyncio, json
+```
+
+``` python
+def user(text): return Msg(role='user', content=[Text(text)])
+
+async def stream(msgs, model, max_think=0, **kw):
+    "Print streamed parts and return the final Completion; limit printed thinking deltas with max_think."
+    cnt = 0
+    async for o in await acomplete(msgs, model, stream=True, **kw):
+        if not isinstance(o, Part): continue
+        if isinstance(o, Thinking):
+            cnt += 1
+            if cnt > max_think: continue
+        print(o.formatted, end='', flush=True)
+    print()
+    return o
+```
+
+``` python
+mtok = 1024
+```
+
+## Chat across providers
+
+Use the same `acomplete` interface with Claude, GPT, Gemini, or Kimi. Set `model` and any provider-specific routing arguments:
+
+``` python
+models = [
+    ('claude-sonnet-5', {}),
+    ('gpt-4o-mini', {}),
+    ('models/gemini-3-flash-preview', {}),
+    ('accounts/fireworks/models/kimi-k3', dict(vendor_name='fireworks_ai'))
+]
+for name, kw in models:
+    r = await acomplete([user("Translate 'hello' into French. Return only the translation.")],
+                       model=name, max_tokens=mtok, **kw)
+    text = ''.join(p.text for p in r.message.content if isinstance(p, Text))
+    print(f"{name}: {text.strip()}")
+```
+
+    claude-sonnet-5: Bonjour
+    gpt-4o-mini: Bonjour
+    models/gemini-3-flash-preview: Bonjour
+    accounts/fireworks/models/kimi-k3: Bonjour
+
+## System prompts
+
+Pass `system` to supply a system prompt. `fastllm` translates it to the provider’s field: Anthropic `system`, OpenAI Responses `instructions`, or Gemini `system_instruction`.
+
+``` python
+sys = "You are a pirate chef. Always respond in pirate speak and mention food. Use one short sentence."
+
+print("Claude: ", end='')
+r = await stream([user("What should I do today?")], model='claude-sonnet-5', system=sys, max_tokens=mtok)
+
+print("Gemini: ", end='')
+r = await stream([user("What should I do today?")], model='models/gemini-3-flash-preview', system=sys, max_tokens=mtok)
+```
+
+    Claude: Arrr, chart a course fer the galley and cook up a hearty stew, matey!
+    Gemini: Sharpen yer cutlass and feast on spicy shark stew, ye scurvy dog!
+
+## Tool calling
+
+Define tools in the shared schema. `fastllm` translates it to the provider’s tool format. This example handles a tool request, adds the result to the history, and continues the conversation:
+
+``` python
+tools = [{"type": "function", "function": {
+    "name": "get_weather",
+    "description": "Get current weather for a city",
+    "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
+}}]
+
+msgs = [user("What's the weather in Paris?")]
+r1 = await stream(msgs, model='claude-sonnet-5', tools=tools, max_tokens=mtok)
+print("Tool calls:", r1.tool_calls)
+```
+
+    - ⏳ `get_weather(city="Paris")` ⏳
+
+    Tool calls: [ToolUse(raw={'caller': {'type': 'direct'}}, cache_control=None, id='toolu_01XacBpJHnsjtPWJWr55nA2m', name='get_weather', arguments={'city': 'Paris'}, server=False, text=None)]
+
+``` python
+msgs += [r1.message, mk_tool_res_msg(r1.tool_calls, ['22°C, sunny with light clouds']),
+         user("Should I bring a jacket? Answer in one short sentence.")]
+r2 = await stream(msgs, model='claude-sonnet-5', tools=tools, max_tokens=mtok)
+```
+
+    No, a jacket isn't necessary — it's a mild, sunny 22°C day.
+
+## Tool Choice
+
+Control whether the model must use tools, can’t use tools, or decides on its own:
+
+``` python
+r = await acomplete([user("Hello there!")], model='claude-sonnet-5',
+                    tools=tools, tool_choice='required', max_tokens=mtok)
+print("Required:", [tc.name for tc in r.tool_calls])
+
+r = await acomplete([user("What's the weather?")], model='claude-sonnet-5',
+                    tools=tools, tool_choice='none', max_tokens=mtok)
+print("None:", r.tool_calls)
+```
+
+    Required: ['get_weather']
+    None: []
+
+## Thinking and extended reasoning
+
+Set `reasoning_effort` to `low`, `medium`, or `high` for supported models. The adapters translate these values to provider-specific settings. Reasoning output uses `Thinking` parts; this example prints a short excerpt alongside the answer.
+
+``` python
+print("Claude: ", end='')
+r = await stream([user("What is 127 × 849? Return just the number.")], model='claude-sonnet-4-6',
+                 reasoning_effort='low', max_tokens=8192)
+for p in r.message.content:
+    if isinstance(p, Thinking): print(f"Thinking excerpt: {p.text[:150]}...")
+
+print("Kimi: ", end='')
+r = await stream([user("What is 127 × 849? Return just the number.")],
+                 model='accounts/fireworks/models/kimi-k3', vendor_name='fireworks_ai',
+                 reasoning_effort='low', max_tokens=8192)
+for p in r.message.content:
+    if isinstance(p, Thinking): print(f"Thinking excerpt: {p.text[:150]}...")
+```
+
+    Claude: 107823
+    Thinking excerpt: 127 × 849 = 127 × 800 + 127 × 49 = 101600 + 6223 = 107823...
+    Kimi: 107823
+    Thinking excerpt: 127 × 849 = 107823. Let me verify: 127*800=101600; 127*49=6223; total 107823. Return just number....
+
+## Web Search (Server Tools)
+
+OpenAI’s Responses API supports server-side web search. Server tool calls are normalized alongside regular tool calls:
+
+``` python
+ws_tools = [{"type": "web_search_preview"}]
+r = await acomplete([user("What is the latest Python release? Reply with the version number only.")],
+                    model='gpt-4o-mini', tools=ws_tools, max_tokens=512)
+print("Server tools used:", [tc.name for tc in r.tool_calls if tc.server])
+```
+
+    Server tools used: ['web_search']
+
+## Caching with Anthropic
+
+Set a part’s `cache_control` to request prompt caching. Later requests can reuse eligible cached content. The example prints usage for two calls with the same system prompt:
+
+``` python
+long_ctx = "You are an expert on the solar system. " * 200
+system = Text(long_ctx, cache_control={'type': 'ephemeral'})
+
+r1 = await acomplete([user("What is Jupiter's mass? Answer in one sentence.")],
+                     model='claude-sonnet-5', system=system, max_tokens=mtok)
+print("First call: cache creation tokens", r1.usage.cache_creation_tokens)
+
+r2 = await acomplete([user("What is Saturn's mass? Answer in one sentence.")],
+                     model='claude-sonnet-5', system=system, max_tokens=mtok)
+print("Second call: cache read tokens", r2.usage.cached_tokens)
+```
+
+    First call: cache creation tokens 2204
+    Second call: cache read tokens 2204
+
+## Media inputs
+
+Use `InputImage` with a model that supports images. The example sends the same image message to three providers:
+
+``` python
+img_msg = Msg(role='user', content=[
+    InputImage(sample_img_url),
+    Text("List three visible objects, using nouns only.")
+])
+
+for name, kw in [('claude-sonnet-5', {}), ('gpt-4o-mini', {}), ('models/gemini-3-flash-preview', {})]:
+    print(f"{name}: ", end='')
+    r = await stream([img_msg], model=name, max_tokens=mtok, **kw)
+```
+
+    claude-sonnet-5: Mountains, lake, trees
+    gpt-4o-mini: 1. Mountains  
+    2. Trees  
+    3. Lake
+    models/gemini-3-flash-preview: planks, lake, mountains
+
+The media adapters support the following part types:
+
+| Media part   | Anthropic | OpenAI Responses | OpenAI Chat | Gemini |
+|--------------|-----------|------------------|-------------|--------|
+| `InputImage` | Yes       | Yes              | Yes         | Yes    |
+| `InputAudio` | No        | No               | Yes         | Yes    |
+| `InputVideo` | No        | No               | No          | Yes    |
+| `InputFile`  | Yes       | Yes              | Yes         | Yes    |
+
+Media parts accept a URL or base64 data URL as their `text`. An unsupported combination raises `ValueError`.

{python_fastllm-0.0.58 → python_fastllm-0.0.60}/fastllm/__init__.py

@@ -2,4 +2,4 @@
 
 - `fastllm.responses`: Provider-independent Responses resources and streaming over FastLLM"""
 
-__version__ = "0.0.58"
+__version__ = "0.0.60"

{python_fastllm-0.0.58 → python_fastllm-0.0.60}/fastllm/_modidx.py

@@ -22,6 +22,8 @@ d = { 'settings': { 'branch': 'main',
             'fastllm.anthropic': { 'fastllm.anthropic._add_cache_control': ('anthropic.html#_add_cache_control', 'fastllm/anthropic.py'),
                                    '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._clear_cache_control': ( 'anthropic.html#_clear_cache_control',
+                                                                               'fastllm/anthropic.py'),
                                    'fastllm.anthropic._sanid': ('anthropic.html#_sanid', 'fastllm/anthropic.py'),
                                    'fastllm.anthropic._url_cite': ('anthropic.html#_url_cite', 'fastllm/anthropic.py'),
                                    'fastllm.anthropic.acollect_stream': ('anthropic.html#acollect_stream', 'fastllm/anthropic.py'),
@@ -140,7 +142,8 @@ d = { 'settings': { 'branch': 'main',
                                'fastllm.codex.main': ('codex.html#main', 'fastllm/codex.py'),
                                'fastllm.codex.mk_jwt': ('codex.html#mk_jwt', 'fastllm/codex.py'),
                                'fastllm.codex.token_expired': ('codex.html#token_expired', 'fastllm/codex.py')},
-            'fastllm.gemini': { 'fastllm.gemini._fn_calls': ('gemini.html#_fn_calls', 'fastllm/gemini.py'),
+            'fastllm.gemini': { 'fastllm.gemini._fn_call': ('gemini.html#_fn_call', 'fastllm/gemini.py'),
+                                'fastllm.gemini._fn_calls': ('gemini.html#_fn_calls', 'fastllm/gemini.py'),
                                 'fastllm.gemini._gem_filter_sch': ('gemini.html#_gem_filter_sch', 'fastllm/gemini.py'),
                                 'fastllm.gemini._gem_part_type': ('gemini.html#_gem_part_type', 'fastllm/gemini.py'),
                                 'fastllm.gemini._url_cite': ('gemini.html#_url_cite', 'fastllm/gemini.py'),
@@ -315,7 +318,9 @@ d = { 'settings': { 'branch': 'main',
                                    'fastllm.responses._msg_parts': ('responses.html#_msg_parts', 'fastllm/responses.py'),
                                    'fastllm.responses._new_id': ('responses.html#_new_id', 'fastllm/responses.py'),
                                    'fastllm.responses._response_base': ('responses.html#_response_base', 'fastllm/responses.py'),
+                                   'fastllm.responses._server_call': ('responses.html#_server_call', 'fastllm/responses.py'),
                                    'fastllm.responses._text_content': ('responses.html#_text_content', 'fastllm/responses.py'),
+                                   'fastllm.responses._visible_parts': ('responses.html#_visible_parts', 'fastllm/responses.py'),
                                    'fastllm.responses.normalize_call_ids': ('responses.html#normalize_call_ids', 'fastllm/responses.py'),
                                    'fastllm.responses.pending_response': ('responses.html#pending_response', 'fastllm/responses.py'),
                                    'fastllm.responses.response_event': ('responses.html#response_event', 'fastllm/responses.py'),

{python_fastllm-0.0.58 → python_fastllm-0.0.60}/fastllm/acomplete.py

@@ -1,4 +1,4 @@
-"""A high level unified function make api calls"""
+"""A unified async API for model completions and streaming"""
 
 # AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/06_acomplete.ipynb.
 

{python_fastllm-0.0.58 → python_fastllm-0.0.60}/fastllm/anthropic.py

@@ -1,3 +1,5 @@
+"""Translate Anthropic Messages API requests, responses, and streams to and from fastllm's shared message types, with tool replay, media, and prompt caching"""
+
 # AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/04_anthropic.ipynb.
 
 # %% auto #0
@@ -16,7 +18,7 @@ from fasttransport.errors import api_error_from_event
 
 from .types import *
 from aidialog.msg_parts import (MdStr, Part, PartType, mk_part, Msg, Completion, Text, Thinking, ToolUse, ToolResult, tool_text,
-                                InputImage, InputAudio, InputVideo, InputFile, data_url)
+    InputImage, InputAudio, InputVideo, InputFile, data_url)
 from .streaming import *
 from .streaming import mk_acollect_stream
 
@@ -44,8 +46,8 @@ def norm_usage(resp):
     pt = int(usg.get("input_tokens", 0) or 0) + cached + cache_creation
     ct = int(usg.get("output_tokens", 0) or 0)
     return Usage(prompt_tokens=pt, completion_tokens=ct, total_tokens=pt + ct,
-                 cached_tokens=cached, cache_creation_tokens=cache_creation, reasoning_tokens=0, raw=usg,
-                 search_queries=int(nested_idx(usg, 'server_tool_use', 'web_search_requests') or 0))
+        cached_tokens=cached, cache_creation_tokens=cache_creation, reasoning_tokens=0, raw=usg,
+        search_queries=int(nested_idx(usg, 'server_tool_use', 'web_search_requests') or 0))
 
 def finalize_usage(usg, parts):
     "Split reasoning tokens out of `output_tokens`: the reported `thinking_tokens`, else an estimate from the thinking text."
@@ -110,7 +112,9 @@ def norm_tr_parts(content):
         if typ=='text': parts.append(Text(b.get('text',''), raw=b))
         elif typ=='image':
             s = b['source']
-            parts.append(InputImage(f"data:{s['media_type']};base64,{s['data']}", mime=s['media_type'], raw=b))
+            if s['type']=='url': parts.append(InputImage(s['url'], raw=b))
+            elif s['type']=='base64': parts.append(InputImage(f"data:{s['media_type']};base64,{s['data']}", mime=s['media_type'], raw=b))
+            else: raise ValueError(f"Unsupported Anthropic image source type: {s['type']!r}")
         else: parts.append(mk_part(**b))
     return parts
 
@@ -238,13 +242,11 @@ def denorm_reasoning(v):
 # %% ../nbs/04_anthropic.ipynb #8fa9fbb8
 def denorm_web_search(v):
     "Map canonical web_search_options to Anthropic hosted web_search tool."
-    _max_uses = {"low": 1, "medium": 5, "high": 10}
+    _max_uses = dict(low=1, medium=5, high=10)
     t = {"type": "web_search_20250305", "name": "web_search"}
     if (typ := (v or {}).get("type")): t["type"] = typ
-    if (s := (v or {}).get("search_context_size")):
-        t["max_uses"] = _max_uses.get(s, 5)
-    if (u := (v or {}).get("user_location", {}).get("approximate")):
-        t["user_location"] = {"type": "approximate", **u}
+    if (s := (v or {}).get("search_context_size")): t["max_uses"] = _max_uses.get(s, 5)
+    if (u := (v or {}).get("user_location", {}).get("approximate")): t["user_location"] = {"type": "approximate", **u}
     return t
 
 # %% ../nbs/04_anthropic.ipynb #6b485275
@@ -269,12 +271,12 @@ def denorm_user(m:Msg):
 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": dict(type="base64", media_type=b64[0], data=data)}
     return {"type": "image", "source": {"type": "url", "url": p.text}}
 
 # %% ../nbs/04_anthropic.ipynb #fc6bbdfc
 def denorm_file(p):
-    if (b64:=data_url(p.text)): return {"type": "document", "source": {"type": "base64", "media_type": b64[0], "data": b64[1]}}
+    if (b64:=data_url(p.text)): return {"type": "document", "source": dict(type="base64", media_type=b64[0], data=b64[1])}
     return {"type": "document", "source": {"type": "url", "url": p.text}}
 
 # %% ../nbs/04_anthropic.ipynb #37b63cc2
@@ -303,15 +305,21 @@ def _add_cache_control(
     return msg
 
 # %% ../nbs/04_anthropic.ipynb #6bc52ef5
+def _clear_cache_control(obj):
+    "Clear canonical cache marks from a message or part."
+    if isinstance(obj, Msg):
+        for p in obj.content: _clear_cache_control(p)
+    elif isinstance(obj, Part): obj.cache_control = None
+
 def apply_cache_idxs(
     msgs,            # Canonical Msgs to mark
     cache_idxs=[-1], # Breakpoint idxs into `msgs` (tool-role msgs excluded); 0 is `sp` when given
     sp=None,         # System prompt: str, Msg, or Part
     ttl=None         # Cache TTL: '5m' (default) or '1h'
 ):
-    "Strip all cache breakpoints from `msgs`, then mark `cache_idxs`; returns `sp`, marked if 0 is included"
-    for m in msgs:
-        for p in m.content: p.cache_control = None
+    "Replace cache breakpoints in `msgs` and `sp`, then return the system prompt."
+    for m in msgs: _clear_cache_control(m)
+    _clear_cache_control(sp)
     if sp:
         if 0 in cache_idxs:
             if isinstance(sp, str): sp = Text(sp)
@@ -335,10 +343,8 @@ def mk_payload(msgs, model, **kwargs):
     if tools:=kwargs.get('tools'):          payload['tools'] = denorm_tool_schs(tools)
     if tchc:=kwargs.get('tool_choice'):     payload['tool_choice'] = denorm_tool_choice(tchc)
     if thk:=kwargs.get('reasoning_effort'): payload.update(denorm_reasoning(thk))
-    if (wopts:=kwargs.get('web_search_options')) is not None: 
-        payload.setdefault('tools', []).append(denorm_web_search(wopts))
-    if (temp:=kwargs.get('temperature')) is not None: 
-        payload['temperature'] = temp
+    if (wopts:=kwargs.get('web_search_options')) is not None: payload.setdefault('tools', []).append(denorm_web_search(wopts))
+    if (temp:=kwargs.get('temperature')) is not None: payload['temperature'] = temp
     return payload
 
 # %% ../nbs/04_anthropic.ipynb #60d52c1f
@@ -367,5 +373,4 @@ def cost(usage, m):
 # %% ../nbs/04_anthropic.ipynb #f7c0b989
 api_registry.register('anthropic', norm_tool_calls=norm_tool_calls, norm_parts=norm_parts, norm_finish=norm_finish, norm_usage=norm_usage,
     finalize_usage=finalize_usage, raw_msg=raw_msg, collate_raw=collate_raw, acollect_stream=acollect_stream, mk_payload=mk_payload,
-    cost=cost, get_hdrs=get_hdrs,
-    endpoint='/v1/messages')
+    cost=cost, get_hdrs=get_hdrs, endpoint='/v1/messages')