smart-docgen 0.2.2__tar.gz → 0.2.3__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.
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/CHANGELOG.md +7 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/PKG-INFO +24 -6
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/README.md +23 -5
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/__init__.py +27 -11
- smart_docgen-0.2.3/docpilot/exceptions.py +63 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/claude.py +3 -1
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/gemini.py +3 -1
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/openai.py +3 -1
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/openai_compat.py +56 -4
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/rag.py +21 -2
- smart_docgen-0.2.3/docpilot/tokens.py +11 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/smart_docgen.egg-info/PKG-INFO +24 -6
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/smart_docgen.egg-info/SOURCES.txt +1 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/smart_docgen.egg-info/scm_file_list.json +102 -101
- smart_docgen-0.2.3/smart_docgen.egg-info/scm_version.json +8 -0
- smart_docgen-0.2.2/docpilot/exceptions.py +0 -34
- smart_docgen-0.2.2/smart_docgen.egg-info/scm_version.json +0 -8
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/.claude/skills/docpilot-tools/SKILL.md +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/.claude/skills/docpilot-tools/docpilot_cli.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/.env.example +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/.github/workflows/publish.yml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/.gitignore +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/LICENSE +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/base.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/docx_analyzer.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/docx_builder.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/html_blocks.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/html_table.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwp_convert.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwpx_analyzer.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwpx_auto_builder.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwpx_builder.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwpx_bullets.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwpx_dynamic_builder.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwpx_section_generator.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/hwpx_style_extractor.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/builder/pdf_builder.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/db/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/db/client.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/db/indexer.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/db/schema.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/db/template_store.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/ingestion/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/ingestion/docx.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/ingestion/hwp.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/ingestion/hwpx.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/ingestion/models.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/ingestion/pdf.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/ingestion/text.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/base.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/benchmark.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mapping/sidecar.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/mcp_server.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/_filter.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/alias.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/embedding.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/eval.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/exact.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/highlight.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/hybrid.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/models.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/morpheme.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/search/reranker.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/template_generator/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/template_generator/analyzer.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/template_generator/docx_extractor.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/template_generator/extractor.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/template_generator/generator.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/Contents/content.hpf +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/Contents/header.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/Contents/section0.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/META-INF/container.rdf +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/META-INF/container.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/META-INF/manifest.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/Preview/PrvImage.png +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/Preview/PrvText.txt +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/mimetype +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/settings.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/base/version.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/gonmun/header.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/gonmun/section0.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/gonmun/sidecar.json +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/minutes/header.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/minutes/section0.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/minutes/sidecar.json +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/proposal/header.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/proposal/section0.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/proposal/sidecar.json +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/report/header.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/report/section0.xml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/docpilot/templates/report/sidecar.json +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/example/run.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/pyproject.toml +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/chunk_size_compare.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/gen_/355/225/231/355/232/214/353/263/264/352/263/240/354/204/234.py" +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/gen_/355/225/231/355/232/214/353/263/264/352/263/240/354/204/234_hwpx_from_docx.py" +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/inspect_index.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/make_ebs_template.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/test_docgen.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/test_folder_search.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/test_hwpx_to_docx.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/test_mapping.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/test_ollama_truncation.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/test_quality.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/test_real.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/scripts/verify_html_render.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/setup.cfg +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/smart_docgen.egg-info/dependency_links.txt +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/smart_docgen.egg-info/entry_points.txt +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/smart_docgen.egg-info/requires.txt +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/smart_docgen.egg-info/top_level.txt +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/conftest.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/mocks/__init__.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/mocks/db_mock.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/mocks/llm_mock.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_alias.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_analyzers.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_builder.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_collection.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_embed_batch_bench.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_embed_model_bench.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_embed_quality_bench.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_embedding.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_generate_from_content.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_hwp_ingestion.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_index_perf.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_indexer.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_ingestion.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_mapping.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_mcp_server.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_search.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_search_quality.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/tests/test_template_generator.py +0 -0
- {smart_docgen-0.2.2 → smart_docgen-0.2.3}/uv.lock +0 -0
|
@@ -5,6 +5,13 @@
|
|
|
5
5
|
|
|
6
6
|
## [Unreleased]
|
|
7
7
|
|
|
8
|
+
## [0.2.3] - 2026-07-27
|
|
9
|
+
|
|
10
|
+
### Added
|
|
11
|
+
- `DocPilot(llm="ollama", num_ctx=...)` — Ollama 요청에 `num_ctx`를 실어 보내고, RAG/content로 조립된 프롬프트가 예상 토큰 기준으로 `num_ctx`를 초과하면 API 호출 전에 `ContextExceededError`로 즉시 차단 (기존엔 조용히 잘리거나 응답이 중간에 끊길 수 있었음)
|
|
12
|
+
- `RagMapper.map()`: `ContextExceededError` 발생 시 `top_k`를 절반씩 줄여 최대 2회 자동 재시도 (Ollama 사전 차단 + Claude/OpenAI/Gemini/Grok의 실제 API 거부 에러 양쪽 모두에 적용)
|
|
13
|
+
- Ollama 매퍼 기본 요청 타임아웃 180초 (`OllamaMapper(timeout=...)`로 조정 가능) — VRAM 부족 시 에러 없이 CPU로 전환되며 20~50배 느려지는 상황에 대한 안전장치
|
|
14
|
+
|
|
8
15
|
## [0.2.2] - 2026-07-24
|
|
9
16
|
|
|
10
17
|
### Added
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: smart-docgen
|
|
3
|
-
Version: 0.2.
|
|
3
|
+
Version: 0.2.3
|
|
4
4
|
Summary: LLM-powered document generation from data sources and templates
|
|
5
5
|
License: MIT
|
|
6
6
|
Project-URL: Homepage, https://github.com/pjkwon/docpilot
|
|
@@ -151,11 +151,26 @@ pip install "smart-docgen[pdf,openai,postgres]" # 풀 스택
|
|
|
151
151
|
|
|
152
152
|
**Ollama 등 로컬 모델의 컨텍스트 한도.** Ollama는 요청에 `num_ctx`를 명시하지 않으면 모델이 지원하는
|
|
153
153
|
최대 컨텍스트보다 훨씬 작은 기본값(예: 4096 — 모델이 32k를 지원해도)으로 로드합니다. 실제 로드된 값은
|
|
154
|
-
`ollama ps`로 확인할 수 있습니다.
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
|
|
154
|
+
`ollama ps`로 확인할 수 있습니다. `DocPilot(llm="ollama", num_ctx=8192)`처럼 `num_ctx`를 넘기면
|
|
155
|
+
그 값이 요청마다 Ollama에 전달되고(서버가 그 이상으로 미리 로드돼 있어야 합니다 — `num_ctx`는 요청
|
|
156
|
+
옵션일 뿐 서버가 로드한 컨텍스트 자체를 늘리지 않습니다. 서버 쪽 최대치는 여전히
|
|
157
|
+
`OLLAMA_CONTEXT_LENGTH` 환경변수나 커스텀 Modelfile로 설정), RAG/content로 조립된 프롬프트가
|
|
158
|
+
바이트 기반 추정치로 `num_ctx`(+ `max_tokens`)를 넘으면 API 호출 전에 `ContextExceededError`로
|
|
159
|
+
즉시 막습니다(추정치라 ±30% 오차 가능 — 정확한 카운트는 Ollama가 토큰 카운팅 API를 제공하지 않아
|
|
160
|
+
불가능합니다). `num_ctx`를 넘기지 않으면 이전과 동일하게 아무 사전 검증 없이 그대로 전송되며, Ollama
|
|
161
|
+
기본값에 걸려 프롬프트 앞부분이 조용히 잘리거나 응답이 중간에 끊길 수 있습니다.
|
|
162
|
+
|
|
163
|
+
`generate()`(RAG 경로)를 쓰면 `ContextExceededError`가 나든(Ollama 사전 차단), Claude/OpenAI/Gemini/Grok이
|
|
164
|
+
컨텍스트 초과로 요청을 거부하든 `RagMapper`가 자동으로 `top_k`를 절반씩 줄여 최대 2회 재시도합니다
|
|
165
|
+
(10 → 5 → 2 등, 1까지 줄여도 실패하면 최종적으로 에러를 올립니다). `generate_from_content()`처럼 RAG가
|
|
166
|
+
없는 경로는 `top_k` 자체가 없어 재시도 대상이 아니며, `max_input_tokens`로 사전에 상한을 걸어야 합니다.
|
|
167
|
+
|
|
168
|
+
Ollama는 컨텍스트 초과와 별개로 VRAM이 부족하면 에러 없이 CPU로 스필오버되며 20~50배 느려질 수 있어
|
|
169
|
+
("멈춘 것"처럼 보임), `OllamaMapper`/`DocPilot(llm="ollama", ...)`는 기본 180초 요청 타임아웃을 둡니다
|
|
170
|
+
— 정상적인 요청도 하드웨어에 따라 이보다 오래 걸릴 수 있으니 필요시 `OllamaMapper(timeout=...)`로
|
|
171
|
+
조정하세요. 응답이 max_tokens/컨텍스트 한도로 잘리면 `MappingError` 메시지에 "max_tokens 제한으로 잘림"이
|
|
172
|
+
명시되어 일반 JSON 파싱 실패와 구분되므로, 이 메시지가 뜨면 `top_k`를 줄이거나 `num_ctx`(및 서버의
|
|
173
|
+
컨텍스트 한도)를 늘려보세요.
|
|
159
174
|
|
|
160
175
|
## LLM 제공자
|
|
161
176
|
|
|
@@ -1348,6 +1363,7 @@ from docpilot.exceptions import (
|
|
|
1348
1363
|
DocPilotError,
|
|
1349
1364
|
IngestionError,
|
|
1350
1365
|
MappingError,
|
|
1366
|
+
ContextExceededError, # MappingError의 서브클래스 — 컨텍스트 초과로 거부/차단된 경우
|
|
1351
1367
|
BuilderError,
|
|
1352
1368
|
ConversionError,
|
|
1353
1369
|
SearchError,
|
|
@@ -1356,6 +1372,8 @@ from docpilot.exceptions import (
|
|
|
1356
1372
|
|
|
1357
1373
|
try:
|
|
1358
1374
|
pilot.generate(...)
|
|
1375
|
+
except ContextExceededError as e:
|
|
1376
|
+
print(e.detail) # RagMapper가 top_k를 줄여가며 재시도했지만 끝내 실패한 경우
|
|
1359
1377
|
except MappingError as e:
|
|
1360
1378
|
print(e.detail) # 상세 오류 메시지
|
|
1361
1379
|
except DocPilotError as e:
|
|
@@ -85,11 +85,26 @@ pip install "smart-docgen[pdf,openai,postgres]" # 풀 스택
|
|
|
85
85
|
|
|
86
86
|
**Ollama 등 로컬 모델의 컨텍스트 한도.** Ollama는 요청에 `num_ctx`를 명시하지 않으면 모델이 지원하는
|
|
87
87
|
최대 컨텍스트보다 훨씬 작은 기본값(예: 4096 — 모델이 32k를 지원해도)으로 로드합니다. 실제 로드된 값은
|
|
88
|
-
`ollama ps`로 확인할 수 있습니다.
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
88
|
+
`ollama ps`로 확인할 수 있습니다. `DocPilot(llm="ollama", num_ctx=8192)`처럼 `num_ctx`를 넘기면
|
|
89
|
+
그 값이 요청마다 Ollama에 전달되고(서버가 그 이상으로 미리 로드돼 있어야 합니다 — `num_ctx`는 요청
|
|
90
|
+
옵션일 뿐 서버가 로드한 컨텍스트 자체를 늘리지 않습니다. 서버 쪽 최대치는 여전히
|
|
91
|
+
`OLLAMA_CONTEXT_LENGTH` 환경변수나 커스텀 Modelfile로 설정), RAG/content로 조립된 프롬프트가
|
|
92
|
+
바이트 기반 추정치로 `num_ctx`(+ `max_tokens`)를 넘으면 API 호출 전에 `ContextExceededError`로
|
|
93
|
+
즉시 막습니다(추정치라 ±30% 오차 가능 — 정확한 카운트는 Ollama가 토큰 카운팅 API를 제공하지 않아
|
|
94
|
+
불가능합니다). `num_ctx`를 넘기지 않으면 이전과 동일하게 아무 사전 검증 없이 그대로 전송되며, Ollama
|
|
95
|
+
기본값에 걸려 프롬프트 앞부분이 조용히 잘리거나 응답이 중간에 끊길 수 있습니다.
|
|
96
|
+
|
|
97
|
+
`generate()`(RAG 경로)를 쓰면 `ContextExceededError`가 나든(Ollama 사전 차단), Claude/OpenAI/Gemini/Grok이
|
|
98
|
+
컨텍스트 초과로 요청을 거부하든 `RagMapper`가 자동으로 `top_k`를 절반씩 줄여 최대 2회 재시도합니다
|
|
99
|
+
(10 → 5 → 2 등, 1까지 줄여도 실패하면 최종적으로 에러를 올립니다). `generate_from_content()`처럼 RAG가
|
|
100
|
+
없는 경로는 `top_k` 자체가 없어 재시도 대상이 아니며, `max_input_tokens`로 사전에 상한을 걸어야 합니다.
|
|
101
|
+
|
|
102
|
+
Ollama는 컨텍스트 초과와 별개로 VRAM이 부족하면 에러 없이 CPU로 스필오버되며 20~50배 느려질 수 있어
|
|
103
|
+
("멈춘 것"처럼 보임), `OllamaMapper`/`DocPilot(llm="ollama", ...)`는 기본 180초 요청 타임아웃을 둡니다
|
|
104
|
+
— 정상적인 요청도 하드웨어에 따라 이보다 오래 걸릴 수 있으니 필요시 `OllamaMapper(timeout=...)`로
|
|
105
|
+
조정하세요. 응답이 max_tokens/컨텍스트 한도로 잘리면 `MappingError` 메시지에 "max_tokens 제한으로 잘림"이
|
|
106
|
+
명시되어 일반 JSON 파싱 실패와 구분되므로, 이 메시지가 뜨면 `top_k`를 줄이거나 `num_ctx`(및 서버의
|
|
107
|
+
컨텍스트 한도)를 늘려보세요.
|
|
93
108
|
|
|
94
109
|
## LLM 제공자
|
|
95
110
|
|
|
@@ -1282,6 +1297,7 @@ from docpilot.exceptions import (
|
|
|
1282
1297
|
DocPilotError,
|
|
1283
1298
|
IngestionError,
|
|
1284
1299
|
MappingError,
|
|
1300
|
+
ContextExceededError, # MappingError의 서브클래스 — 컨텍스트 초과로 거부/차단된 경우
|
|
1285
1301
|
BuilderError,
|
|
1286
1302
|
ConversionError,
|
|
1287
1303
|
SearchError,
|
|
@@ -1290,6 +1306,8 @@ from docpilot.exceptions import (
|
|
|
1290
1306
|
|
|
1291
1307
|
try:
|
|
1292
1308
|
pilot.generate(...)
|
|
1309
|
+
except ContextExceededError as e:
|
|
1310
|
+
print(e.detail) # RagMapper가 top_k를 줄여가며 재시도했지만 끝내 실패한 경우
|
|
1293
1311
|
except MappingError as e:
|
|
1294
1312
|
print(e.detail) # 상세 오류 메시지
|
|
1295
1313
|
except DocPilotError as e:
|
|
@@ -175,6 +175,10 @@ def _b(base_url: str | None) -> dict:
|
|
|
175
175
|
return {"base_url": base_url} if base_url else {}
|
|
176
176
|
|
|
177
177
|
|
|
178
|
+
def _n(num_ctx: int | None) -> dict:
|
|
179
|
+
return {"num_ctx": num_ctx} if num_ctx else {}
|
|
180
|
+
|
|
181
|
+
|
|
178
182
|
def _ingest_instructions_doc(path: Path) -> str:
|
|
179
183
|
"""지침 문서를 읽어 텍스트로 반환한다. 지원하지 않는 형식이면 빈 문자열 반환."""
|
|
180
184
|
from docpilot.ingestion import text as text_ing
|
|
@@ -258,12 +262,6 @@ def _estimate_tokens_from_folder(folder: Path, n_sections: int) -> int:
|
|
|
258
262
|
_DEFAULT_MAX_INPUT_TOKENS_WARN = 50_000
|
|
259
263
|
|
|
260
264
|
|
|
261
|
-
def _estimate_tokens_raw(text: str) -> int:
|
|
262
|
-
"""Rough token estimate for text that goes into the prompt in full — no RAG retrieval
|
|
263
|
-
subset, so (unlike _estimate_tokens_from_folder) no per-section cap applies."""
|
|
264
|
-
return int(len(text.encode("utf-8")) * 0.4)
|
|
265
|
-
|
|
266
|
-
|
|
267
265
|
def _resolve_instructions(
|
|
268
266
|
extra_instructions: str | None,
|
|
269
267
|
sidecar_instructions: str,
|
|
@@ -313,7 +311,8 @@ def _resolve_content(content: str | list[str | Path] | list[IngestedDocument]) -
|
|
|
313
311
|
def _check_content_size(content_str: str, max_input_tokens: int | None) -> None:
|
|
314
312
|
"""Warn or raise on a rough byte-based token estimate — *_from_content() methods send
|
|
315
313
|
content in full with no RAG top_k to bound its size automatically."""
|
|
316
|
-
|
|
314
|
+
from docpilot.tokens import estimate_tokens_raw
|
|
315
|
+
estimated_tokens = estimate_tokens_raw(content_str)
|
|
317
316
|
if max_input_tokens is not None:
|
|
318
317
|
if estimated_tokens > max_input_tokens:
|
|
319
318
|
raise MappingError(
|
|
@@ -1022,17 +1021,27 @@ class DocPilot:
|
|
|
1022
1021
|
api_key: str | None = None,
|
|
1023
1022
|
model: str | None = None,
|
|
1024
1023
|
base_url: str | None = None,
|
|
1024
|
+
num_ctx: int | None = None,
|
|
1025
1025
|
database_url: str | None = None,
|
|
1026
1026
|
embed_fn=None,
|
|
1027
1027
|
use_reranker: bool = False,
|
|
1028
1028
|
) -> None:
|
|
1029
|
+
"""
|
|
1030
|
+
num_ctx: Ollama-only — the context window (input+output token budget) the server
|
|
1031
|
+
loaded the model with. Ollama defaults to a small value (often 4096) regardless
|
|
1032
|
+
of what the model actually supports, and silently truncates prompts that exceed
|
|
1033
|
+
it rather than erroring — see README "Ollama 등 로컬 모델의 컨텍스트 한도". Pass the
|
|
1034
|
+
same value your Ollama server was started/configured with (check via `ollama ps`)
|
|
1035
|
+
so docpilot can (a) forward it as a per-request option and (b) reject oversized
|
|
1036
|
+
RAG/content prompts before sending them, instead of silently losing context.
|
|
1037
|
+
"""
|
|
1029
1038
|
self._llm = llm or os.environ.get("DOCPILOT_LLM", "claude")
|
|
1030
1039
|
self._api_key = api_key
|
|
1031
1040
|
if embed_fn is None:
|
|
1032
1041
|
from docpilot.search.embedding import default_embed_fn
|
|
1033
1042
|
embed_fn = default_embed_fn() # None when sentence-transformers not installed
|
|
1034
1043
|
self._embed_fn = embed_fn
|
|
1035
|
-
base_mapper = self._build_mapper(self._llm, api_key, model, base_url)
|
|
1044
|
+
base_mapper = self._build_mapper(self._llm, api_key, model, base_url, num_ctx)
|
|
1036
1045
|
|
|
1037
1046
|
from docpilot.mapping.rag import RagMapper
|
|
1038
1047
|
self._mapper = base_mapper
|
|
@@ -1459,7 +1468,8 @@ class DocPilot:
|
|
|
1459
1468
|
est_output = len(sections) * _EST_OUTPUT_TOKENS_PER_SECTION
|
|
1460
1469
|
|
|
1461
1470
|
if quick:
|
|
1462
|
-
|
|
1471
|
+
from docpilot.tokens import estimate_tokens_raw
|
|
1472
|
+
input_tokens = estimate_tokens_raw(content_str)
|
|
1463
1473
|
input_cost = input_tokens / 1_000_000 * in_price
|
|
1464
1474
|
output_cost = est_output / 1_000_000 * out_price
|
|
1465
1475
|
lines = [
|
|
@@ -1605,7 +1615,13 @@ class DocPilot:
|
|
|
1605
1615
|
return suggest_extras(folder)
|
|
1606
1616
|
|
|
1607
1617
|
@staticmethod
|
|
1608
|
-
def _build_mapper(
|
|
1618
|
+
def _build_mapper(
|
|
1619
|
+
llm: str,
|
|
1620
|
+
api_key: str | None,
|
|
1621
|
+
model: str | None,
|
|
1622
|
+
base_url: str | None,
|
|
1623
|
+
num_ctx: int | None = None,
|
|
1624
|
+
):
|
|
1609
1625
|
from docpilot.mapping import ClaudeMapper, OpenAIMapper, GeminiMapper
|
|
1610
1626
|
from docpilot.mapping.openai_compat import GrokMapper, OllamaMapper
|
|
1611
1627
|
|
|
@@ -1619,7 +1635,7 @@ class DocPilot:
|
|
|
1619
1635
|
case "grok":
|
|
1620
1636
|
return GrokMapper(api_key=api_key, **(_m(model)))
|
|
1621
1637
|
case "ollama":
|
|
1622
|
-
return OllamaMapper(**(_m(model)), **(_b(base_url)))
|
|
1638
|
+
return OllamaMapper(**(_m(model)), **(_b(base_url)), **(_n(num_ctx)))
|
|
1623
1639
|
case _:
|
|
1624
1640
|
raise MappingError(
|
|
1625
1641
|
f"Unknown LLM '{llm}'. "
|
|
@@ -0,0 +1,63 @@
|
|
|
1
|
+
from __future__ import annotations
|
|
2
|
+
|
|
3
|
+
|
|
4
|
+
class DocPilotError(Exception):
|
|
5
|
+
"""Base exception for all docpilot errors."""
|
|
6
|
+
|
|
7
|
+
def __init__(self, message: str, detail: str | None = None) -> None:
|
|
8
|
+
self.detail = detail
|
|
9
|
+
full_message = f"{message} — {detail}" if detail else message
|
|
10
|
+
super().__init__(full_message)
|
|
11
|
+
|
|
12
|
+
|
|
13
|
+
class IngestionError(DocPilotError):
|
|
14
|
+
"""Raised when a source file cannot be parsed or read."""
|
|
15
|
+
|
|
16
|
+
|
|
17
|
+
class MappingError(DocPilotError):
|
|
18
|
+
"""Raised when LLM fails to map content to template sections."""
|
|
19
|
+
|
|
20
|
+
|
|
21
|
+
class ContextExceededError(MappingError):
|
|
22
|
+
"""Raised when a request would exceed (Ollama: local pre-flight estimate) or was
|
|
23
|
+
rejected for exceeding (hosted APIs: live error from the provider) the model's
|
|
24
|
+
context window. Kept distinct from MappingError so callers like RagMapper can retry
|
|
25
|
+
with a smaller top_k specifically for this failure, without swallowing unrelated
|
|
26
|
+
errors (auth, rate limit, etc.) that a retry can't fix."""
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
_CONTEXT_OVERFLOW_MARKERS = (
|
|
30
|
+
"context length",
|
|
31
|
+
"context_length_exceeded",
|
|
32
|
+
"maximum context",
|
|
33
|
+
"context window",
|
|
34
|
+
"too many tokens",
|
|
35
|
+
"exceeds the maximum number of tokens",
|
|
36
|
+
"token limit",
|
|
37
|
+
)
|
|
38
|
+
|
|
39
|
+
|
|
40
|
+
def is_context_overflow_error(message: str) -> bool:
|
|
41
|
+
"""Heuristic match against a provider SDK error message to tell a context-window
|
|
42
|
+
rejection apart from other API failures. Providers don't share a common exception
|
|
43
|
+
type for this, so this is a substring match, not a guarantee — false negatives just
|
|
44
|
+
fall through to the generic MappingError path (no retry); false positives cost at
|
|
45
|
+
most a couple of wasted retries before the error surfaces anyway."""
|
|
46
|
+
lowered = message.lower()
|
|
47
|
+
return any(marker in lowered for marker in _CONTEXT_OVERFLOW_MARKERS)
|
|
48
|
+
|
|
49
|
+
|
|
50
|
+
class BuilderError(DocPilotError):
|
|
51
|
+
"""Raised when document generation fails."""
|
|
52
|
+
|
|
53
|
+
|
|
54
|
+
class ConversionError(DocPilotError):
|
|
55
|
+
"""Raised when a file format conversion fails."""
|
|
56
|
+
|
|
57
|
+
|
|
58
|
+
class SearchError(DocPilotError):
|
|
59
|
+
"""Raised when a search or index operation fails."""
|
|
60
|
+
|
|
61
|
+
|
|
62
|
+
class TemplateError(DocPilotError):
|
|
63
|
+
"""Raised when template analysis or generation fails."""
|
|
@@ -4,7 +4,7 @@ import logging
|
|
|
4
4
|
import os
|
|
5
5
|
import time
|
|
6
6
|
|
|
7
|
-
from docpilot.exceptions import MappingError
|
|
7
|
+
from docpilot.exceptions import ContextExceededError, MappingError, is_context_overflow_error
|
|
8
8
|
from docpilot.mapping.base import BaseLLMMapper, MappingResult, TemplateSection
|
|
9
9
|
|
|
10
10
|
logger = logging.getLogger(__name__)
|
|
@@ -61,6 +61,8 @@ class ClaudeMapper(BaseLLMMapper):
|
|
|
61
61
|
messages=[{"role": "user", "content": prompt}],
|
|
62
62
|
)
|
|
63
63
|
except Exception as e:
|
|
64
|
+
if is_context_overflow_error(str(e)):
|
|
65
|
+
raise ContextExceededError("Claude API가 컨텍스트 초과로 요청을 거부함", detail=str(e)) from e
|
|
64
66
|
raise MappingError("Claude API call failed", detail=str(e)) from e
|
|
65
67
|
elapsed = time.perf_counter() - start
|
|
66
68
|
|
|
@@ -4,7 +4,7 @@ import logging
|
|
|
4
4
|
import os
|
|
5
5
|
import time
|
|
6
6
|
|
|
7
|
-
from docpilot.exceptions import MappingError
|
|
7
|
+
from docpilot.exceptions import ContextExceededError, MappingError, is_context_overflow_error
|
|
8
8
|
from docpilot.mapping.base import BaseLLMMapper, MappingResult, TemplateSection
|
|
9
9
|
|
|
10
10
|
logger = logging.getLogger(__name__)
|
|
@@ -68,6 +68,8 @@ class GeminiMapper(BaseLLMMapper):
|
|
|
68
68
|
config=types.GenerateContentConfig(max_output_tokens=self._max_tokens),
|
|
69
69
|
)
|
|
70
70
|
except Exception as e:
|
|
71
|
+
if is_context_overflow_error(str(e)):
|
|
72
|
+
raise ContextExceededError("Gemini API가 컨텍스트 초과로 요청을 거부함", detail=str(e)) from e
|
|
71
73
|
raise MappingError("Gemini API call failed", detail=str(e)) from e
|
|
72
74
|
elapsed = time.perf_counter() - start
|
|
73
75
|
|
|
@@ -4,7 +4,7 @@ import logging
|
|
|
4
4
|
import os
|
|
5
5
|
import time
|
|
6
6
|
|
|
7
|
-
from docpilot.exceptions import MappingError
|
|
7
|
+
from docpilot.exceptions import ContextExceededError, MappingError, is_context_overflow_error
|
|
8
8
|
from docpilot.mapping.base import BaseLLMMapper, MappingResult, TemplateSection
|
|
9
9
|
|
|
10
10
|
logger = logging.getLogger(__name__)
|
|
@@ -66,6 +66,8 @@ class OpenAIMapper(BaseLLMMapper):
|
|
|
66
66
|
messages=[{"role": "user", "content": prompt}],
|
|
67
67
|
)
|
|
68
68
|
except Exception as e:
|
|
69
|
+
if is_context_overflow_error(str(e)):
|
|
70
|
+
raise ContextExceededError("OpenAI API가 컨텍스트 초과로 요청을 거부함", detail=str(e)) from e
|
|
69
71
|
raise MappingError("OpenAI API call failed", detail=str(e)) from e
|
|
70
72
|
elapsed = time.perf_counter() - start
|
|
71
73
|
|
|
@@ -4,8 +4,9 @@ import logging
|
|
|
4
4
|
import os
|
|
5
5
|
import time
|
|
6
6
|
|
|
7
|
-
from docpilot.exceptions import MappingError
|
|
7
|
+
from docpilot.exceptions import ContextExceededError, MappingError, is_context_overflow_error
|
|
8
8
|
from docpilot.mapping.base import BaseLLMMapper, MappingResult, TemplateSection
|
|
9
|
+
from docpilot.tokens import estimate_tokens_raw
|
|
9
10
|
|
|
10
11
|
logger = logging.getLogger(__name__)
|
|
11
12
|
|
|
@@ -25,25 +26,59 @@ class OpenAICompatMapper(BaseLLMMapper):
|
|
|
25
26
|
base_url: str,
|
|
26
27
|
api_key: str = "not-required",
|
|
27
28
|
max_tokens: int = 8096,
|
|
29
|
+
num_ctx: int | None = None,
|
|
30
|
+
timeout: float | None = None,
|
|
28
31
|
) -> None:
|
|
29
32
|
self._model = model
|
|
30
33
|
self._base_url = base_url
|
|
31
34
|
self._api_key = api_key
|
|
32
35
|
self._max_tokens = max_tokens
|
|
36
|
+
self._num_ctx = num_ctx
|
|
37
|
+
self._timeout = timeout
|
|
38
|
+
|
|
39
|
+
def _extra_body(self) -> dict:
|
|
40
|
+
# Ollama-specific request option — openai SDK has no native num_ctx param, so it
|
|
41
|
+
# must be passed through extra_body's "options" dict (Ollama's own convention).
|
|
42
|
+
return {"options": {"num_ctx": self._num_ctx}} if self._num_ctx else {}
|
|
43
|
+
|
|
44
|
+
def _check_num_ctx(self, prompt: str, max_tokens: int) -> None:
|
|
45
|
+
"""Reject prompts that won't fit num_ctx *before* sending — Ollama itself doesn't
|
|
46
|
+
error on this, it silently truncates the prompt from the front instead, which loses
|
|
47
|
+
RAG context with no signal in the response (see README "Ollama 등 로컬 모델의 컨텍스트
|
|
48
|
+
한도"). Estimate only — no exact tokenizer available for Ollama models over the API."""
|
|
49
|
+
if self._num_ctx is None:
|
|
50
|
+
return
|
|
51
|
+
estimated = estimate_tokens_raw(prompt)
|
|
52
|
+
if estimated + max_tokens > self._num_ctx:
|
|
53
|
+
raise ContextExceededError(
|
|
54
|
+
f"예상 입력 토큰({estimated:,}) + max_tokens({max_tokens:,})이 "
|
|
55
|
+
f"num_ctx({self._num_ctx:,})를 초과합니다.",
|
|
56
|
+
detail=(
|
|
57
|
+
"이 상태로 보내면 Ollama가 프롬프트 앞부분을 조용히 잘라내거나 "
|
|
58
|
+
"응답이 중간에 끊길 수 있습니다. num_ctx를 늘리거나 top_k/컨텍스트 "
|
|
59
|
+
"크기를 줄이세요. (추정치는 바이트 기반 휴리스틱으로 ±30% 오차 가능)"
|
|
60
|
+
),
|
|
61
|
+
)
|
|
33
62
|
|
|
34
63
|
def complete(self, prompt: str, max_tokens: int = 2048) -> str:
|
|
35
64
|
try:
|
|
36
65
|
from openai import OpenAI
|
|
37
66
|
except ImportError as e:
|
|
38
67
|
raise MappingError("openai SDK required: pip install openai") from e
|
|
39
|
-
|
|
68
|
+
self._check_num_ctx(prompt, max_tokens)
|
|
69
|
+
client = OpenAI(api_key=self._api_key, base_url=self._base_url, timeout=self._timeout)
|
|
40
70
|
try:
|
|
41
71
|
response = client.chat.completions.create(
|
|
42
72
|
model=self._model,
|
|
43
73
|
max_tokens=max_tokens,
|
|
44
74
|
messages=[{"role": "user", "content": prompt}],
|
|
75
|
+
extra_body=self._extra_body(),
|
|
45
76
|
)
|
|
46
77
|
except Exception as e:
|
|
78
|
+
if is_context_overflow_error(str(e)):
|
|
79
|
+
raise ContextExceededError(
|
|
80
|
+
f"API가 컨텍스트 초과로 요청을 거부함 ({self._base_url})", detail=str(e)
|
|
81
|
+
) from e
|
|
47
82
|
raise MappingError(f"API call failed ({self._base_url})", detail=str(e)) from e
|
|
48
83
|
return response.choices[0].message.content or ""
|
|
49
84
|
|
|
@@ -58,8 +93,9 @@ class OpenAICompatMapper(BaseLLMMapper):
|
|
|
58
93
|
except ImportError as e:
|
|
59
94
|
raise MappingError("openai SDK required: pip install openai") from e
|
|
60
95
|
|
|
61
|
-
client = OpenAI(api_key=self._api_key, base_url=self._base_url)
|
|
96
|
+
client = OpenAI(api_key=self._api_key, base_url=self._base_url, timeout=self._timeout)
|
|
62
97
|
prompt = self._build_prompt(self._resolve_content(content), sections, instructions)
|
|
98
|
+
self._check_num_ctx(prompt, self._max_tokens)
|
|
63
99
|
|
|
64
100
|
start = time.perf_counter()
|
|
65
101
|
try:
|
|
@@ -67,8 +103,13 @@ class OpenAICompatMapper(BaseLLMMapper):
|
|
|
67
103
|
model=self._model,
|
|
68
104
|
max_tokens=self._max_tokens,
|
|
69
105
|
messages=[{"role": "user", "content": prompt}],
|
|
106
|
+
extra_body=self._extra_body(),
|
|
70
107
|
)
|
|
71
108
|
except Exception as e:
|
|
109
|
+
if is_context_overflow_error(str(e)):
|
|
110
|
+
raise ContextExceededError(
|
|
111
|
+
f"API가 컨텍스트 초과로 요청을 거부함 ({self._base_url})", detail=str(e)
|
|
112
|
+
) from e
|
|
72
113
|
raise MappingError(
|
|
73
114
|
f"API call failed ({self._base_url})", detail=str(e)
|
|
74
115
|
) from e
|
|
@@ -110,10 +151,21 @@ def GrokMapper(
|
|
|
110
151
|
return OpenAICompatMapper(model=model, base_url=_GROK_BASE_URL, api_key=key, max_tokens=max_tokens)
|
|
111
152
|
|
|
112
153
|
|
|
154
|
+
_OLLAMA_DEFAULT_TIMEOUT = 180.0 # seconds — Ollama has no explicit context-overflow error;
|
|
155
|
+
# a VRAM-starved request degrades to CPU offload instead of failing (20-50x slower, can look
|
|
156
|
+
# hung), so this is a starting-point safety net rather than a measured value. Pass timeout=
|
|
157
|
+
# to OllamaMapper()/OpenAICompatMapper() directly to tune for your hardware.
|
|
158
|
+
|
|
159
|
+
|
|
113
160
|
def OllamaMapper(
|
|
114
161
|
model: str = "llama3.2",
|
|
115
162
|
base_url: str | None = None,
|
|
116
163
|
max_tokens: int = 8096,
|
|
164
|
+
num_ctx: int | None = None,
|
|
165
|
+
timeout: float = _OLLAMA_DEFAULT_TIMEOUT,
|
|
117
166
|
) -> OpenAICompatMapper:
|
|
118
167
|
url = base_url or os.environ.get("OLLAMA_BASE_URL", _OLLAMA_BASE_URL)
|
|
119
|
-
return OpenAICompatMapper(
|
|
168
|
+
return OpenAICompatMapper(
|
|
169
|
+
model=model, base_url=url, api_key="ollama", max_tokens=max_tokens,
|
|
170
|
+
num_ctx=num_ctx, timeout=timeout,
|
|
171
|
+
)
|
|
@@ -1,16 +1,22 @@
|
|
|
1
1
|
from __future__ import annotations
|
|
2
2
|
|
|
3
|
+
import logging
|
|
3
4
|
from collections.abc import Callable
|
|
4
5
|
from typing import TYPE_CHECKING
|
|
5
6
|
|
|
7
|
+
from docpilot.exceptions import ContextExceededError
|
|
6
8
|
from docpilot.mapping.base import BaseLLMMapper, MappingResult, TemplateSection
|
|
7
9
|
from docpilot.search.models import SearchFilter, SearchResult
|
|
8
10
|
|
|
9
11
|
if TYPE_CHECKING:
|
|
10
12
|
from docpilot.ingestion.models import IngestedDocument
|
|
11
13
|
|
|
14
|
+
logger = logging.getLogger(__name__)
|
|
15
|
+
|
|
12
16
|
EmbedFn = Callable[[str], list[float]]
|
|
13
17
|
|
|
18
|
+
_MAX_CONTEXT_RETRIES = 2
|
|
19
|
+
|
|
14
20
|
|
|
15
21
|
class RagMapper:
|
|
16
22
|
"""
|
|
@@ -48,8 +54,21 @@ class RagMapper:
|
|
|
48
54
|
top_k: int | None = None,
|
|
49
55
|
filters: SearchFilter | None = None,
|
|
50
56
|
) -> MappingResult:
|
|
51
|
-
|
|
52
|
-
|
|
57
|
+
effective_top_k = top_k if top_k is not None else self._top_k
|
|
58
|
+
|
|
59
|
+
for attempt in range(_MAX_CONTEXT_RETRIES + 1):
|
|
60
|
+
content = self.retrieve_content(sections, top_k=effective_top_k, filters=filters)
|
|
61
|
+
try:
|
|
62
|
+
return self._mapper.map(content, sections, instructions)
|
|
63
|
+
except ContextExceededError:
|
|
64
|
+
if attempt >= _MAX_CONTEXT_RETRIES or effective_top_k <= 1:
|
|
65
|
+
raise
|
|
66
|
+
new_top_k = max(1, effective_top_k // 2)
|
|
67
|
+
logger.warning(
|
|
68
|
+
"컨텍스트 초과로 top_k %d → %d로 줄여 재시도 (%d/%d)",
|
|
69
|
+
effective_top_k, new_top_k, attempt + 1, _MAX_CONTEXT_RETRIES,
|
|
70
|
+
)
|
|
71
|
+
effective_top_k = new_top_k
|
|
53
72
|
|
|
54
73
|
def retrieve_content(
|
|
55
74
|
self,
|
|
@@ -0,0 +1,11 @@
|
|
|
1
|
+
from __future__ import annotations
|
|
2
|
+
|
|
3
|
+
_TOKENS_PER_BYTE = 0.4
|
|
4
|
+
|
|
5
|
+
|
|
6
|
+
def estimate_tokens_raw(text: str) -> int:
|
|
7
|
+
"""Rough token estimate for text sent to an LLM in full — ~0.4 tokens/byte, tuned for
|
|
8
|
+
Korean/mixed text. No tokenizer call, so this is a heuristic (±30%), not an exact count;
|
|
9
|
+
exact counts require the model's own tokenizer, which local models (Ollama) don't expose
|
|
10
|
+
over the API."""
|
|
11
|
+
return int(len(text.encode("utf-8")) * _TOKENS_PER_BYTE)
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: smart-docgen
|
|
3
|
-
Version: 0.2.
|
|
3
|
+
Version: 0.2.3
|
|
4
4
|
Summary: LLM-powered document generation from data sources and templates
|
|
5
5
|
License: MIT
|
|
6
6
|
Project-URL: Homepage, https://github.com/pjkwon/docpilot
|
|
@@ -151,11 +151,26 @@ pip install "smart-docgen[pdf,openai,postgres]" # 풀 스택
|
|
|
151
151
|
|
|
152
152
|
**Ollama 등 로컬 모델의 컨텍스트 한도.** Ollama는 요청에 `num_ctx`를 명시하지 않으면 모델이 지원하는
|
|
153
153
|
최대 컨텍스트보다 훨씬 작은 기본값(예: 4096 — 모델이 32k를 지원해도)으로 로드합니다. 실제 로드된 값은
|
|
154
|
-
`ollama ps`로 확인할 수 있습니다.
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
|
|
154
|
+
`ollama ps`로 확인할 수 있습니다. `DocPilot(llm="ollama", num_ctx=8192)`처럼 `num_ctx`를 넘기면
|
|
155
|
+
그 값이 요청마다 Ollama에 전달되고(서버가 그 이상으로 미리 로드돼 있어야 합니다 — `num_ctx`는 요청
|
|
156
|
+
옵션일 뿐 서버가 로드한 컨텍스트 자체를 늘리지 않습니다. 서버 쪽 최대치는 여전히
|
|
157
|
+
`OLLAMA_CONTEXT_LENGTH` 환경변수나 커스텀 Modelfile로 설정), RAG/content로 조립된 프롬프트가
|
|
158
|
+
바이트 기반 추정치로 `num_ctx`(+ `max_tokens`)를 넘으면 API 호출 전에 `ContextExceededError`로
|
|
159
|
+
즉시 막습니다(추정치라 ±30% 오차 가능 — 정확한 카운트는 Ollama가 토큰 카운팅 API를 제공하지 않아
|
|
160
|
+
불가능합니다). `num_ctx`를 넘기지 않으면 이전과 동일하게 아무 사전 검증 없이 그대로 전송되며, Ollama
|
|
161
|
+
기본값에 걸려 프롬프트 앞부분이 조용히 잘리거나 응답이 중간에 끊길 수 있습니다.
|
|
162
|
+
|
|
163
|
+
`generate()`(RAG 경로)를 쓰면 `ContextExceededError`가 나든(Ollama 사전 차단), Claude/OpenAI/Gemini/Grok이
|
|
164
|
+
컨텍스트 초과로 요청을 거부하든 `RagMapper`가 자동으로 `top_k`를 절반씩 줄여 최대 2회 재시도합니다
|
|
165
|
+
(10 → 5 → 2 등, 1까지 줄여도 실패하면 최종적으로 에러를 올립니다). `generate_from_content()`처럼 RAG가
|
|
166
|
+
없는 경로는 `top_k` 자체가 없어 재시도 대상이 아니며, `max_input_tokens`로 사전에 상한을 걸어야 합니다.
|
|
167
|
+
|
|
168
|
+
Ollama는 컨텍스트 초과와 별개로 VRAM이 부족하면 에러 없이 CPU로 스필오버되며 20~50배 느려질 수 있어
|
|
169
|
+
("멈춘 것"처럼 보임), `OllamaMapper`/`DocPilot(llm="ollama", ...)`는 기본 180초 요청 타임아웃을 둡니다
|
|
170
|
+
— 정상적인 요청도 하드웨어에 따라 이보다 오래 걸릴 수 있으니 필요시 `OllamaMapper(timeout=...)`로
|
|
171
|
+
조정하세요. 응답이 max_tokens/컨텍스트 한도로 잘리면 `MappingError` 메시지에 "max_tokens 제한으로 잘림"이
|
|
172
|
+
명시되어 일반 JSON 파싱 실패와 구분되므로, 이 메시지가 뜨면 `top_k`를 줄이거나 `num_ctx`(및 서버의
|
|
173
|
+
컨텍스트 한도)를 늘려보세요.
|
|
159
174
|
|
|
160
175
|
## LLM 제공자
|
|
161
176
|
|
|
@@ -1348,6 +1363,7 @@ from docpilot.exceptions import (
|
|
|
1348
1363
|
DocPilotError,
|
|
1349
1364
|
IngestionError,
|
|
1350
1365
|
MappingError,
|
|
1366
|
+
ContextExceededError, # MappingError의 서브클래스 — 컨텍스트 초과로 거부/차단된 경우
|
|
1351
1367
|
BuilderError,
|
|
1352
1368
|
ConversionError,
|
|
1353
1369
|
SearchError,
|
|
@@ -1356,6 +1372,8 @@ from docpilot.exceptions import (
|
|
|
1356
1372
|
|
|
1357
1373
|
try:
|
|
1358
1374
|
pilot.generate(...)
|
|
1375
|
+
except ContextExceededError as e:
|
|
1376
|
+
print(e.detail) # RagMapper가 top_k를 줄여가며 재시도했지만 끝내 실패한 경우
|
|
1359
1377
|
except MappingError as e:
|
|
1360
1378
|
print(e.detail) # 상세 오류 메시지
|
|
1361
1379
|
except DocPilotError as e:
|