Coverage for src/lilbee/core/config/enums.py: 100%

1"""StrEnum types used by :mod:`lilbee.config`.""" 

2 

3from enum import StrEnum 

4 

5 

6class ChatMode(StrEnum): 

7 """How chat turns route through retrieval. ``search`` uses retrieval; ``chat`` skips it.""" 

8 

9 SEARCH = "search" 

10 CHAT = "chat" 

11 

12 

13class LlmProvider(StrEnum): 

14 """Inference backend that ``create_provider`` builds. 

15 

16 ``auto`` prefix-routes: native GGUF refs to the local llama-server engine, 

17 remote-prefixed refs (``ollama/``, ``openai/``, ...) to the SDK backend. 

18 ``remote`` forces the SDK backend. 

19 """ 

20 

21 AUTO = "auto" 

22 REMOTE = "remote" 

23 

24 

25class RerankerType(StrEnum): 

26 """How the reranker GGUF is served. ``auto`` detects by architecture.""" 

27 

28 AUTO = "auto" 

29 CROSS_ENCODER = "cross_encoder" 

30 LLM = "llm" 

31 

32 

33class CrawlRenderMode(StrEnum): 

34 """How a crawl fetches pages. ``http`` uses no browser; ``browser`` runs Chromium with JS.""" 

35 

36 HTTP = "http" 

37 BROWSER = "browser" 

38 

39 

40class ClustererBackend(StrEnum): 

41 """Known wiki clusterer backends.""" 

42 

43 EMBEDDING = "embedding" 

44 CONCEPTS = "concepts" 

45 

46 

47class WikiEntityMode(StrEnum): 

48 """Strategy used to extract entities for the wiki. 

49 

50 The extractor emits typed NER entities only. Concept pages are 

51 proposed by the LLM inside the per-source batched call in 

52 :mod:`lilbee.wiki.generation`. The enum values reflect that 

53 extractor responsibility. 

54 """ 

55 

56 NER_ENTITIES = "ner_entities" 

57 NER_CONCEPTS_PLUS_LLM_TYPES = "ner_concepts_plus_llm_types" 

58 LLM_TAGGED = "llm_tagged" 

59 

60 

61class TableModel(StrEnum): 

62 """xberg's table structure recognition model, used when layout detection is on. 

63 

64 The ``slanet_*`` variants are the docling-parity lineage; ``tatr`` is xberg's 

65 older default. ``disabled`` skips structure recognition. 

66 """ 

67 

68 DISABLED = "disabled" 

69 TATR = "tatr" 

70 SLANET_AUTO = "slanet_auto" 

71 SLANET_PLUS = "slanet_plus" 

72 SLANET_WIRED = "slanet_wired" 

73 SLANET_WIRELESS = "slanet_wireless" 

74 

75 

76class KvCacheType(StrEnum): 

77 """KV cache element type. ``q8_0`` / ``q4_0`` require flash attention.""" 

78 

79 F16 = "f16" 

80 F32 = "f32" 

81 Q8_0 = "q8_0" 

82 Q4_0 = "q4_0" 

83 

84 

85# Bytes per KV element for memory budgeting. The quantized variants are 

86# ~1 byte of data plus shared scales, close enough for context-fit math. 

87KV_CACHE_TYPE_BYTES: dict[KvCacheType, int] = { 

88 KvCacheType.F16: 2, 

89 KvCacheType.F32: 4, 

90 KvCacheType.Q8_0: 1, 

91 KvCacheType.Q4_0: 1, 

92}