caesar-agent 0.3.14__tar.gz → 0.4.28__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (152) hide show
  1. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/LICENSE +26 -0
  2. caesar_agent-0.4.28/PKG-INFO +410 -0
  3. caesar_agent-0.4.28/README.md +335 -0
  4. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/analysis/llm_as_judge.py +1 -1
  5. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/analysis/transfer_configs.py +1 -1
  6. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/artifact_synthesis.py +370 -126
  7. caesar_agent-0.4.28/caesar/arxiv_explorer.py +394 -0
  8. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/brave_search.py +58 -7
  9. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/caesar_agent.py +74 -39
  10. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/caesar_config.py +111 -8
  11. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_agi/improve_transformer.yaml +2 -2
  12. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_agi/intelligence_bottleneck.yaml +2 -2
  13. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_agi/single_agent_agi.yaml +2 -2
  14. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_agi/solve_arcagi.yaml +2 -2
  15. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_creative/constrained_creativity.yaml +3 -3
  16. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_creative/counterfactual_reasoning.yaml +3 -3
  17. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_creative/crossdomain_synthesis.yaml +3 -3
  18. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_creative/meta_creativity.yaml +3 -3
  19. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_creative/openended_creativity.yaml +3 -3
  20. caesar_agent-0.4.28/caesar/config/config_preset/arxiv.yaml +81 -0
  21. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_preset/mini.yaml +3 -3
  22. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_preset/nano.yaml +2 -2
  23. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_preset/regular.yaml +9 -5
  24. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_test/single_agent_test.yaml +2 -2
  25. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_wikipedia/single_agent_ai.yaml +2 -2
  26. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_wikipedia/single_agent_evo.yaml +2 -2
  27. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_wikipedia/single_agent_sci.yaml +2 -2
  28. caesar_agent-0.4.28/caesar/config/llm_as_judge/llm_prompts/nus_rubric_4pt_api.txt +87 -0
  29. caesar_agent-0.4.28/caesar/image_generator.py +1720 -0
  30. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/agg5.py +27 -0
  31. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/build5.py +42 -0
  32. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/cmp.py +46 -0
  33. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/compare.py +70 -0
  34. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/diag.py +55 -0
  35. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/summ.py +57 -0
  36. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/tally.py +38 -0
  37. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/v3an.py +57 -0
  38. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/v4an.py +57 -0
  39. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/v5an.py +72 -0
  40. caesar_agent-0.4.28/caesar/query_result/1_10_all_agent_answers/nus_rubric_study/analysis/why.py +39 -0
  41. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/run_agent.py +52 -5
  42. caesar_agent-0.4.28/caesar/semantic_scholar.py +207 -0
  43. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/web_explorer.py +165 -13
  44. caesar_agent-0.4.28/caesar_agent.egg-info/PKG-INFO +410 -0
  45. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar_agent.egg-info/SOURCES.txt +23 -1
  46. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar_agent.egg-info/requires.txt +16 -8
  47. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/pyproject.toml +21 -10
  48. caesar_agent-0.4.28/requirements.txt +64 -0
  49. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/agent_memory.py +39 -5
  50. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/base_agent.py +28 -5
  51. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/config.py +5 -2
  52. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/kb_client.py +177 -106
  53. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/kb_server.py +112 -26
  54. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/llm_handler.py +237 -27
  55. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/logger.py +17 -13
  56. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/process.py +24 -2
  57. caesar_agent-0.4.28/test/test_arxiv_explorer.py +439 -0
  58. caesar_agent-0.4.28/test/test_arxiv_integration_manual.py +188 -0
  59. caesar_agent-0.4.28/test/test_kb_rerank_fallback.py +87 -0
  60. caesar_agent-0.4.28/test/test_public_mode_key.py +90 -0
  61. caesar_agent-0.4.28/test/test_reasoning_model_gating.py +72 -0
  62. caesar_agent-0.4.28/test/test_url_guard.py +261 -0
  63. caesar_agent-0.3.14/PKG-INFO +0 -278
  64. caesar_agent-0.3.14/README.md +0 -208
  65. caesar_agent-0.3.14/caesar_agent.egg-info/PKG-INFO +0 -278
  66. caesar_agent-0.3.14/requirements.txt +0 -45
  67. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/NOTICE +0 -0
  68. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/__init__.py +0 -0
  69. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/analysis/judge_analysis.py +0 -0
  70. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/analysis/prepare_artifact.py +0 -0
  71. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/checkpoint.py +0 -0
  72. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_test/batch_test.jsonl +0 -0
  73. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/config_test/batch_test.yaml +0 -0
  74. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/custom_role/self_transcend_insights_v1.txt +0 -0
  75. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/custom_role/self_transcend_insights_v2.txt +0 -0
  76. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/custom_role/self_transcend_insights_v3.txt +0 -0
  77. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/custom_role/self_transcend_insights_v4.txt +0 -0
  78. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/custom_role/self_transcend_role_v1.txt +0 -0
  79. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/custom_role/self_transcend_role_v2.txt +0 -0
  80. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/custom_role/self_transcend_role_v3.txt +0 -0
  81. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/agent_names/agent_names_all.txt +0 -0
  82. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/agent_names/agent_names_caesar.txt +0 -0
  83. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/human_prompts/human_eval_prompt.txt +0 -0
  84. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/human_prompts/human_eval_prompt2.txt +0 -0
  85. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/eli5_prompt_long.txt +0 -0
  86. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/eli5_prompt_medium.txt +0 -0
  87. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/eli5_prompt_short.txt +0 -0
  88. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/judge_csv_prompt.txt +0 -0
  89. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/judge_summary_prompt.txt +0 -0
  90. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/nus_rubric_10pt_api.txt +0 -0
  91. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/nus_rubric_10pt_new.txt +0 -0
  92. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/nus_rubric_10pt_old.txt +0 -0
  93. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/query_prompt_deep.txt +0 -0
  94. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/llm_prompts/query_prompt_shallow.txt +0 -0
  95. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/queries/queries_n50.json +0 -0
  96. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/queries/queries_n50.jsonl +0 -0
  97. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_bat_ca1.txt +0 -0
  98. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_bear_ca3.txt +0 -0
  99. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_cat_cam.txt +0 -0
  100. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_cow_cld.txt +0 -0
  101. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_dog_cls.txt +0 -0
  102. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_fox_ged.txt +0 -0
  103. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_lion_ges.txt +0 -0
  104. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_owl_gpd.txt +0 -0
  105. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/answer_pig_gps.txt +0 -0
  106. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/judge_claude.txt +0 -0
  107. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/judge_gemini.txt +0 -0
  108. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/judge_gpt.txt +0 -0
  109. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar/config/llm_as_judge/templates/query.txt +0 -0
  110. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar_agent.egg-info/dependency_links.txt +0 -0
  111. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar_agent.egg-info/entry_points.txt +0 -0
  112. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/caesar_agent.egg-info/top_level.txt +0 -0
  113. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/__init__.py +0 -0
  114. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/__init__.py +0 -0
  115. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/action.py +0 -0
  116. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/advanced_reset_action.py +0 -0
  117. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/edit_code_action.py +0 -0
  118. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/edit_test_action.py +0 -0
  119. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/exec_code_action.py +0 -0
  120. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/priority_search_action.py +0 -0
  121. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/reset_action.py +0 -0
  122. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/revert_code_action.py +0 -0
  123. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/save_kb_action.py +0 -0
  124. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/tournament_search_action.py +0 -0
  125. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action/transition_action.py +0 -0
  126. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/action_selector.py +0 -0
  127. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/agent.py +0 -0
  128. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/agent_api.py +0 -0
  129. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/database.py +0 -0
  130. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/executor.py +0 -0
  131. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/fsm.py +0 -0
  132. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/fsm_selector.py +0 -0
  133. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/history.py +0 -0
  134. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/metadata.py +0 -0
  135. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/multi_agent.py +0 -0
  136. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/parsing.py +0 -0
  137. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/repository.py +0 -0
  138. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/rome/state.py +0 -0
  139. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/setup.cfg +0 -0
  140. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_compressor.py +0 -0
  141. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_database.py +0 -0
  142. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_evalplus_agent.py +0 -0
  143. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_executor.py +0 -0
  144. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_fsm.py +0 -0
  145. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_knowledge_base.py +0 -0
  146. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_llm_handler.py +0 -0
  147. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_logger.py +0 -0
  148. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_metadata.py +0 -0
  149. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_multi_agent.py +0 -0
  150. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_parsing.py +0 -0
  151. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_search.py +0 -0
  152. {caesar_agent-0.3.14 → caesar_agent-0.4.28}/test/test_single_agent.py +0 -0
@@ -1,3 +1,4 @@
1
+
1
2
  Apache License
2
3
  Version 2.0, January 2004
3
4
  http://www.apache.org/licenses/
@@ -174,3 +175,28 @@
174
175
  of your accepting any such warranty or additional liability.
175
176
 
176
177
  END OF TERMS AND CONDITIONS
178
+
179
+ APPENDIX: How to apply the Apache License to your work.
180
+
181
+ To apply the Apache License to your work, attach the following
182
+ boilerplate notice, with the fields enclosed by brackets "[]"
183
+ replaced with your own identifying information. (Don't include
184
+ the brackets!) The text should be enclosed in the appropriate
185
+ comment syntax for the file format. We also recommend that a
186
+ file or class name and description of purpose be included on the
187
+ same "printed page" as the copyright notice for easier
188
+ identification within third-party archives.
189
+
190
+ Copyright [yyyy] [name of copyright owner]
191
+
192
+ Licensed under the Apache License, Version 2.0 (the "License");
193
+ you may not use this file except in compliance with the License.
194
+ You may obtain a copy of the License at
195
+
196
+ http://www.apache.org/licenses/LICENSE-2.0
197
+
198
+ Unless required by applicable law or agreed to in writing, software
199
+ distributed under the License is distributed on an "AS IS" BASIS,
200
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
201
+ See the License for the specific language governing permissions and
202
+ limitations under the License.
@@ -0,0 +1,410 @@
1
+ Metadata-Version: 2.4
2
+ Name: caesar-agent
3
+ Version: 0.4.28
4
+ Summary: Caesar: autonomous AI research agent with graph-based deep web exploration and adversarial answer synthesis.
5
+ Author: Jason Liang, Elliot Meyerson, Risto Miikkulainen
6
+ Maintainer-email: Jason Liang <jasonzliang@utexas.edu>
7
+ License-Expression: Apache-2.0
8
+ Project-URL: Homepage, https://github.com/jasonzliang/caesar-agent
9
+ Project-URL: Repository, https://github.com/jasonzliang/caesar-agent
10
+ Project-URL: Issues, https://github.com/jasonzliang/caesar-agent/issues
11
+ Project-URL: Paper, https://arxiv.org/abs/2604.20855
12
+ Project-URL: Changelog, https://github.com/jasonzliang/caesar-agent/releases
13
+ Keywords: ai-agents,llm,rag,agentic-rag,agentic-ai,autonomous-agents,deep-research,research-agent,knowledge-graph,retrieval-augmented-generation,scientific-discovery,web-exploration,adversarial-refinement,llm-agents,openai,anthropic,chromadb,llama-index,ai-research
14
+ Classifier: Development Status :: 4 - Beta
15
+ Classifier: Intended Audience :: Developers
16
+ Classifier: Intended Audience :: Science/Research
17
+ Classifier: Programming Language :: Python :: 3.10
18
+ Classifier: Programming Language :: Python :: 3.11
19
+ Classifier: Programming Language :: Python :: 3.12
20
+ Classifier: Programming Language :: Python :: 3.13
21
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
22
+ Classifier: Topic :: Scientific/Engineering :: Information Analysis
23
+ Classifier: Topic :: Internet :: WWW/HTTP :: Indexing/Search
24
+ Requires-Python: <3.14,>=3.10
25
+ Description-Content-Type: text/markdown
26
+ License-File: LICENSE
27
+ License-File: NOTICE
28
+ Requires-Dist: litellm
29
+ Requires-Dist: networkx
30
+ Requires-Dist: numpy
31
+ Requires-Dist: openai
32
+ Requires-Dist: portalocker
33
+ Requires-Dist: psutil
34
+ Requires-Dist: pytest
35
+ Requires-Dist: scipy
36
+ Requires-Dist: tinydb
37
+ Requires-Dist: chromadb
38
+ Requires-Dist: llama-index
39
+ Requires-Dist: llama-index-vector-stores-chroma
40
+ Requires-Dist: llama-index-embeddings-openai
41
+ Requires-Dist: llama-index-llms-openai
42
+ Requires-Dist: pillow
43
+ Requires-Dist: mem0ai
44
+ Requires-Dist: anthropic
45
+ Requires-Dist: beautifulsoup4
46
+ Requires-Dist: curl_cffi>=0.16.3
47
+ Requires-Dist: ddgs
48
+ Requires-Dist: google-genai>=2.16.0
49
+ Requires-Dist: pygraphviz
50
+ Requires-Dist: pypdf
51
+ Requires-Dist: pyyaml>=6.0
52
+ Requires-Dist: tenacity
53
+ Requires-Dist: trafilatura>=2.1.0
54
+ Requires-Dist: fastapi
55
+ Requires-Dist: pytz
56
+ Requires-Dist: requests
57
+ Requires-Dist: rich
58
+ Requires-Dist: graphviz
59
+ Requires-Dist: tiktoken
60
+ Requires-Dist: xxhash
61
+ Provides-Extra: compress
62
+ Requires-Dist: llmlingua; extra == "compress"
63
+ Provides-Extra: benchmark
64
+ Requires-Dist: evalplus; extra == "benchmark"
65
+ Provides-Extra: dev
66
+ Requires-Dist: pytest-cov>=4.1.0; extra == "dev"
67
+ Requires-Dist: black>=23.0.0; extra == "dev"
68
+ Requires-Dist: mypy>=1.7.0; extra == "dev"
69
+ Requires-Dist: ipython>=8.0.0; extra == "dev"
70
+ Provides-Extra: test
71
+ Requires-Dist: pytest>=7.4.0; extra == "test"
72
+ Requires-Dist: pytest-asyncio>=0.21.0; extra == "test"
73
+ Requires-Dist: pytest-xdist>=3.5.0; extra == "test"
74
+ Dynamic: license-file
75
+
76
+ <p align="center">
77
+ <img src="https://jasonzliang.github.io/caesar-agent/caesar.webp" alt="Caesar autonomous research agent architecture: Perceive-Think-Act exploration loop and adversarial artifact synthesis" width="720"/>
78
+ </p>
79
+
80
+ <h1 align="center">Caesar: Autonomous AI Research Agent</h1>
81
+
82
+ <p align="center">
83
+ <strong>The open-source alternative to ChatGPT Deep Research and Perplexity.</strong>
84
+ </p>
85
+
86
+ <p align="center">
87
+ <a href="https://caesar.evolution.ml"><strong>Try the live demo</strong></a>
88
+ &nbsp;·&nbsp;
89
+ <a href="https://jasonzliang.github.io/caesar-agent"><strong>Project home page</strong></a>
90
+ </p>
91
+
92
+ <p align="center">
93
+ <a href="https://caesar.evolution.ml"><img alt="Live demo" src="https://img.shields.io/badge/Live%20Demo-Try%20Caesar-2ea44f?logo=googlechrome&logoColor=white"></a>
94
+ <a href="https://arxiv.org/abs/2604.20855"><img alt="arXiv" src="https://img.shields.io/badge/arXiv-2604.20855-b31b1b?logo=arxiv&logoColor=white"></a>
95
+ <a href="https://www.researchgate.net/publication/402554537_Caesar_Deep_Agentic_Web_Exploration_for_Creative_Answer_Synthesis"><img alt="ResearchGate" src="https://img.shields.io/badge/ResearchGate-Caesar-00ccbb?logo=researchgate&logoColor=white"></a>
96
+ <a href="LICENSE"><img alt="License" src="https://img.shields.io/badge/License-Apache%202.0-blue?logo=apache&logoColor=white"></a>
97
+ <a href="https://github.com/jasonzliang/caesar-agent/commits/main"><img alt="Last commit" src="https://img.shields.io/github/last-commit/jasonzliang/caesar-agent?logo=github&logoColor=white&label=Last%20Commit&color=181717"></a>
98
+ </p>
99
+
100
+ **Caesar** is an autonomous AI research agent that navigates the web, reasons
101
+ over a dynamic knowledge graph, and synthesizes novel, grounded answers. In
102
+ blinded LLM-as-a-Judge creativity evaluations, Caesar scored **26.96 / 30** on
103
+ the headline full-answers configuration, beating the runner-up (Gemini 3 Deep
104
+ Research, 23.78) by **3.18 points** and outscoring GPT-5.2 Deep Research (15.74)
105
+ by over **11 points**. Cliff's δ effect sizes are uniformly large (**≥ 0.76**,
106
+ well above the 0.47 large-effect threshold), and the result is corroborated by
107
+ an independent 23-rater human study.
108
+
109
+ If you're looking for an **agentic RAG system that goes beyond retrieval**
110
+ (graph-based exploration, adversarial verification, and multi-draft synthesis),
111
+ this is it.
112
+
113
+ > 📄 **Read the paper:**
114
+ > [_Caesar: Deep Agentic Web Exploration for Creative Answer Synthesis_](https://arxiv.org/abs/2604.20855)
115
+ > (Liang, Meyerson, Miikkulainen, 2026 — **v3**, 8 May 2026) ·
116
+ > [DOI: 10.48550/arXiv.2604.20855](https://doi.org/10.48550/arXiv.2604.20855) ·
117
+ > [PDF](https://arxiv.org/pdf/2604.20855v3)
118
+
119
+ ## Quickstart
120
+
121
+ **From PyPI:**
122
+
123
+ ```bash
124
+ pip install caesar-agent
125
+ export OPENAI_API_KEY=your_key
126
+ caesar regular -q "your research question"
127
+ ```
128
+
129
+ **From source:**
130
+
131
+ ```bash
132
+ git clone https://github.com/jasonzliang/caesar-agent.git
133
+ cd caesar-agent && pip install -e .
134
+ export OPENAI_API_KEY=your_key
135
+ caesar regular -q "your research question"
136
+ ```
137
+
138
+ **In a browser:**
139
+
140
+ ```bash
141
+ cd web_server && ./launch.sh
142
+ # then open http://localhost:3000
143
+ ```
144
+
145
+ The **[Caesar Web Server](web_server/README.md)** is a FastAPI + Next.js GUI
146
+ that submits runs, streams progress, and renders the live knowledge graph and
147
+ final artifact.
148
+
149
+ ## Setup Notes
150
+
151
+ Both install paths give you the same `caesar` console script, so every command
152
+ in this README behaves identically whether you came from PyPI or from a
153
+ checkout. From a checkout you can also skip the script entirely and run
154
+ `python caesar/run_agent.py regular -q "..."` — that still needs the
155
+ dependencies (`pip install -r requirements.txt`), just not the package itself.
156
+
157
+ Everything else worth knowing before your first run:
158
+
159
+ - **Python** — 3.10 through 3.13
160
+ - **Presets** — `nano` (fast, ~$0.80), `mini` (balanced, ~$2), `regular` (deep,
161
+ ~$5–$10). Cost scales with synthesis output tokens, so treat these as
162
+ order-of-magnitude rather than quotes.
163
+ - **API keys** — `OPENAI_API_KEY` alone is enough to run. Add
164
+ `ANTHROPIC_API_KEY` or `GOOGLE_API_KEY` to reach Claude and Gemini models, and
165
+ `BRAVE_API_KEY` for better web search — without it, search falls back to
166
+ DuckDuckGo.
167
+ - **Results** — `caesar/result/` from a source checkout, `~/.caesar/result/`
168
+ from a PyPI install. `CAESAR_RESULT_DIR` overrides both.
169
+ - **Custom configs** — drop a YAML in `~/.caesar/configs/` and call it by name
170
+ (`caesar my_preset -q "..."`). Bundled preset names take precedence, so avoid
171
+ naming yours `nano`, `mini`, or `regular`.
172
+ - **Going deeper** — the full env-var list, exploration modes, and synthesis
173
+ options are in the **[Caesar module docs](caesar/README.md)**.
174
+
175
+ ## What It's Good For
176
+
177
+ Caesar shines on **open-ended, creative, cross-disciplinary** research, where
178
+ retrieval alone won't work:
179
+
180
+ - **Hypothesis generation**: novel cross-domain connections (e.g., bridging
181
+ materials science and biology)
182
+ - **Literature synthesis**: graph-grounded review that spots tensions and gaps
183
+ between papers
184
+ - **arXiv citation-graph review**: traverse the Semantic Scholar citation graph
185
+ of arXiv papers (papers as nodes, citations as edges) instead of the open web
186
+ (`caesar arxiv -q "..."`)
187
+ - **Competitive intelligence**: deep mapping of a technical or market landscape
188
+ - **Counterfactual & meta-creative reasoning**: "what if X was different?" style
189
+ inquiry
190
+ - **Novel solution ideation**: e.g., ARC-AGI–style problem exploration
191
+
192
+ It's **not** the right tool for quick factual lookups or latency-sensitive apps.
193
+ Caesar is designed for depth, not speed.
194
+
195
+ ## Why Caesar?
196
+
197
+ Current deep-research agents (ChatGPT Deep Research, Perplexity, GPT Researcher,
198
+ Gemini Deep Research) optimize for **retrieval precision over a flat sequence of
199
+ documents**. They produce competent summaries but suffer from **navigational
200
+ amnesia**, fall into local minima, and generate derivative, consensus-driven
201
+ outputs.
202
+
203
+ Caesar is different:
204
+
205
+ | Design choice | Caesar | ChatGPT / Gemini Deep Research | Perplexity Research | GPT Researcher |
206
+ | --------------------------------------------------- | :----------------: | :----------------------------: | :-----------------: | :---------------------: |
207
+ | Persistent knowledge graph built during exploration | ✅ | not published | not published | ❌ |
208
+ | Adversarial refinement over its own drafts | ✅ | not published | 🟡 | 🟡 |
209
+ | Serialized graph and JSON run log you keep | ✅ | 🟡 | 🟡 | 🟡 |
210
+ | Runs on your own keys and hardware | ✅ | ❌ | ❌ | ✅ |
211
+ | Works with no setup and no API keys | ❌ | ✅ | ✅ | ❌ |
212
+ | Typical time to an answer | 10 min – 1.5 hrs | 5–30 min | under 3 min | 2–5 min |
213
+ | Browser and mobile access | ❌ | ✅ | ✅ | 🟡 |
214
+ | Maturity | research prototype | GA product | GA product | established open source |
215
+
216
+ <sub>Compared as of July 2026 against each product's then-current tier. **"Not
217
+ published"** means the vendor has not documented the mechanism, which is not the
218
+ same as its absence: these are closed systems and we can report only what they
219
+ disclose. ChatGPT and Deep Research are trademarks of OpenAI; Perplexity of
220
+ Perplexity AI; Gemini of Google.</sub>
221
+
222
+ ## How It Works
223
+
224
+ Caesar operates in two cognitive phases:
225
+
226
+ ### 1. Deep Web Exploration: stateful graph traversal
227
+
228
+ A recursive **Perceive–Think–Act** loop performs topological traversal of
229
+ information spaces. Rather than isolating summaries, Caesar generates
230
+ context-aware insights conditioned on the **local structure of the exploration
231
+ graph**, analyzing how new content builds upon or contradicts neighboring nodes.
232
+ A dynamic policy, informed by a vector knowledge base and episodic memory,
233
+ autonomously switches between depth-first expansion, strategic backtracking, and
234
+ targeted web search.
235
+
236
+ ### 2. Adversarial Artifact Synthesis
237
+
238
+ Rather than a single-pass summary, Caesar runs as a recursive self-correction
239
+ environment. Between drafts the agent re-reads its own artifact and formulates a
240
+ refined, **orthogonal challenge** targeting narrative weaknesses, gaps and
241
+ contradictions in the current belief state. Drafts are produced recurrently,
242
+ each conditioned on the last, then merged — forcing the agent out of the
243
+ consensus basin that traps single-pass LLMs.
244
+
245
+ ## Architectural Innovations
246
+
247
+ - **Domain-Specific Role Adaptation**: the agent rewrites its own system prompt
248
+ per task, overriding the safety-biased generic responses typical of RLHF
249
+ models.
250
+ - **Graph-Augmented Insight Generation**: insights are conditioned on the
251
+ exploration graph neighborhood, enabling online associative reasoning.
252
+ - **Knowledge-Guided Exploration Policy**: detects navigational stagnation via
253
+ episodic memory and forces backtracking.
254
+ - **Adversarial Query Refinement**: orthogonal queries push the agent out of
255
+ generic LLM consensus toward novel, grounded facts.
256
+
257
+ ## Benchmark Results
258
+
259
+ Evaluated with a blinded **3-model LLM-as-a-Judge panel** (Claude Sonnet 4.5,
260
+ GPT-5.2, Gemini 3 Pro) across three creativity dimensions (**New**, **Useful**,
261
+ **Surprising**), scored 1–10 each. Headline configuration: full answers,
262
+ unconstrained length.
263
+
264
+ | Agent | New | Useful | Surprising | **Total** | Cliff's δ |
265
+ | ------------------------ | :------: | :------: | :--------: | :-------: | :-------: |
266
+ | **Caesar** | **9.11** | **8.87** | **8.98** | **26.96** | — |
267
+ | Gemini 3 Deep Research | 8.09 | 7.60 | 8.09 | 23.78 | 0.84 |
268
+ | Sonnet 4.5 Deep Research | 6.73 | 7.49 | 6.42 | 20.64 | 1.00 |
269
+ | GPT-5.2 Deep Research | 5.07 | 6.31 | 4.36 | 15.74 | 1.00 |
270
+
271
+ A **13–23% improvement over state-of-the-art deep research agents**, with Caesar
272
+ leading across all three output formats (Full, ELI5, ELI5-450W). Cliff's δ ≥
273
+ 0.76 in every comparison (well above the 0.47 large-effect threshold); δ = 1.00
274
+ vs. GPT-5.2 Deep Research denotes strict dominance in every output format; vs.
275
+ Sonnet 4.5 Deep Research it is 1.00 on full answers and 0.76–0.96 on the ELI5
276
+ formats. The paper frames these results by magnitude of difference rather than
277
+ null-hypothesis testing: with n = 5 challenges per group, δ estimates stochastic
278
+ dominance and no p-values are reported.
279
+
280
+ **Compute-controlled comparison.** At a matched ~$5/challenge budget (T=250 with
281
+ GPT-5-mini), Caesar still wins: 26.16 vs. Gemini 3 Deep 24.37, Sonnet 4.5 Deep
282
+ 21.00, GPT-5.2 Deep 16.16. The lead is not an artifact of larger compute.
283
+
284
+ **Human evaluation.** 23 raters preferred Caesar in 63 of 112 pairwise A/B
285
+ matchups vs. Gemini 3 Deep Research (56.25%, odds ratio 1.29), independently
286
+ corroborating the LLM-judge findings.
287
+
288
+ Ablations isolate what each component contributes. Removing the knowledge graph
289
+ costs δ = 0.52 (large), and restricting traversal to a single hop costs the
290
+ same; shallow exploration costs δ = 0.92 on total score. Adversarial refinement
291
+ is a trade rather than a uniform gain — it lifts Surprising (+1.76) and New at
292
+ the expense of Useful, and the generative merge recovers utility (7.04 → 8.28)
293
+ while keeping the discovered insights. See the
294
+ [paper](https://arxiv.org/abs/2604.20855) for full methodology,
295
+ per-output-format tables, exploration-budget ablation, and judge bias analysis.
296
+
297
+ ## Example Output
298
+
299
+ After a run, Caesar writes a full artifact (abstract + body with citations) plus
300
+ a structured summary:
301
+
302
+ ```json
303
+ {
304
+ "wall_time": 591.31,
305
+ "tokens_used": 109873,
306
+ "token_cost": 0.29,
307
+ "api_calls": 20,
308
+ "webpages_visited": 4,
309
+ "iterations_elapsed": 5,
310
+ "artifact_dir": "result/.../agent_CaesarExplorer.synthesis.04161850",
311
+ "num_drafts": 2,
312
+ "config_summary": { "...": "..." }
313
+ }
314
+ ```
315
+
316
+ The `artifact_dir` contains one `.txt` per synthesis draft, a final merged
317
+ artifact, and a metadata file tracking sources cited in each draft. Knowledge
318
+ graphs are saved as compressed JSON checkpoints for reproducibility or post-hoc
319
+ analysis.
320
+
321
+ ## Built on Rome
322
+
323
+ Caesar is built on **Rome**, a Finite State Machine framework for stateful AI
324
+ agents that provides the agent runtime (FSM lifecycles, action selection,
325
+ memory, LLM handling) that Caesar builds on, episodic memory, dynamic policy
326
+ routing, and verifiable code execution. See the
327
+ [Rome framework docs](rome/README.md) if you want to build your own agent on
328
+ top.
329
+
330
+ ## Project Layout
331
+
332
+ ```
333
+ caesar-agent/
334
+ ├── caesar/ # Caesar agent (see caesar/README.md for full usage)
335
+ │ ├── caesar_agent.py
336
+ │ ├── artifact_synthesis.py
337
+ │ ├── run_agent.py
338
+ │ ├── config/ # YAML configs and creativity benchmarks
339
+ │ └── paper/ # Caesar paper (PDF)
340
+ ├── rome/ # Rome framework (see rome/README.md): FSM, memory, LLM handlers, KB client
341
+ ├── web_server/ # FastAPI + Next.js web GUI (see web_server/README.md)
342
+ └── web_app/ # Streamlit operator tools: human eval + graph explorer (see web_app/README.md)
343
+ ```
344
+
345
+ ## FAQ
346
+
347
+ **How is this different from LangGraph / CrewAI / AutoGen?** Those are
348
+ orchestration frameworks: they help you wire up agents. Rome is an opinionated
349
+ runtime for _how_ agents should reason (graph-structured exploration,
350
+ adversarial verification, episodic memory). Caesar is a concrete research agent
351
+ built on it.
352
+
353
+ **Do I need GPUs?** No. Caesar uses hosted LLM APIs (OpenAI, Anthropic). A local
354
+ ChromaDB instance handles the vector store. Runs on a laptop.
355
+
356
+ **Which models are supported?** OpenAI (GPT-5 family, o-series reasoning
357
+ models), Anthropic (Claude 4.5 / 4.6), Google (Gemini 3.x), and any
358
+ OpenAI-compatible endpoint. Model selection is per-subsystem (exploration,
359
+ synthesis, judging) via YAML config.
360
+
361
+ **How much does a typical run cost?** A 5-iteration exploration with
362
+ GPT-5.4-mini (`caesar/config/config_test/single_agent_test.yaml`) runs at
363
+ roughly $0.30 and 10 minutes. A 250-iteration deep run with the `regular` preset
364
+ (`gpt-5.6-terra` exploration, `gpt-5.6-sol` synthesis) is typically $5–$10.
365
+
366
+ **Can I reproduce the benchmarks?** Yes. Configs, judge rubrics, and evaluation
367
+ scripts are in `caesar/config/` and `caesar/analysis/`.
368
+
369
+ ## Contributing & Community
370
+
371
+ - ⭐ **Star the repo** if Caesar is useful for your research
372
+ - 💬
373
+ **[Open a Discussion](https://github.com/jasonzliang/caesar-agent/discussions)**
374
+ for ideas, questions, or use cases
375
+ - 🐛 **[File an Issue](https://github.com/jasonzliang/caesar-agent/issues)** for
376
+ bugs or feature requests
377
+ - 🔧 **PRs welcome**, especially new exploration policies, synthesis strategies,
378
+ and benchmark domains
379
+
380
+ ### Good places to start a fork
381
+
382
+ | Goal | Effort |
383
+ | ------------------------------------------------------------------ | :-----: |
384
+ | Adapt Caesar for your research domain (legal, biology, finance, …) | 1–2 hrs |
385
+ | Add a new web-search backend (Tavily, Exa, Serper, …) | 2–3 hrs |
386
+ | Experiment with multi-agent synthesis (ring or debate merge) | 4–6 hrs |
387
+
388
+ If you fork Caesar for your own work,
389
+ [open a Discussion](https://github.com/jasonzliang/caesar-agent/discussions) —
390
+ we'd love to see what you build.
391
+
392
+ ## Citation
393
+
394
+ If you use Caesar in your research, please cite:
395
+
396
+ ```bibtex
397
+ @misc{liang26caesar,
398
+ title={Caesar: Deep Agentic Web Exploration for Creative Answer Synthesis},
399
+ author={Jason Liang and Elliot Meyerson and Risto Miikkulainen},
400
+ year={2026},
401
+ eprint={2604.20855},
402
+ archivePrefix={arXiv},
403
+ primaryClass={cs.IR},
404
+ url={https://arxiv.org/abs/2604.20855},
405
+ }
406
+ ```
407
+
408
+ ## License
409
+
410
+ Apache License 2.0. See [LICENSE](LICENSE).