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 ReasoningMode(StrEnum): 

7 """How a chat API surface presents a reasoning model's thinking. 

8 

9 ``separate`` reports thinking in the wire format's own reasoning channel: 

10 ``reasoning_content`` on ``/v1/chat/completions``, a ``thinking`` block on 

11 ``/v1/messages``. ``inline`` presents thinking as ordinary answer text with 

12 the ``<think>`` markers stripped, for clients that never render the 

13 reasoning channel. ``off`` asks the model not to think, which only reaches 

14 the template on a locally served model; ``/v1/messages`` also drops 

15 thinking a model produces anyway, ``/v1/chat/completions`` still reports it 

16 in ``reasoning_content``. 

17 """ 

18 

19 SEPARATE = "separate" 

20 INLINE = "inline" 

21 OFF = "off" 

22 

23 

24class ChatMode(StrEnum): 

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

26 

27 SEARCH = "search" 

28 CHAT = "chat" 

29 

30 

31class LlmProvider(StrEnum): 

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

33 

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

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

36 ``remote`` forces the SDK backend. 

37 """ 

38 

39 AUTO = "auto" 

40 REMOTE = "remote" 

41 

42 

43class RerankerType(StrEnum): 

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

45 

46 AUTO = "auto" 

47 CROSS_ENCODER = "cross_encoder" 

48 LLM = "llm" 

49 

50 

51class CrawlRenderMode(StrEnum): 

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

53 

54 HTTP = "http" 

55 BROWSER = "browser" 

56 

57 

58class ClustererBackend(StrEnum): 

59 """Known wiki clusterer backends.""" 

60 

61 EMBEDDING = "embedding" 

62 CONCEPTS = "concepts" 

63 

64 

65class WikiEntityMode(StrEnum): 

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

67 

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

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

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

71 extractor responsibility. 

72 """ 

73 

74 NER_ENTITIES = "ner_entities" 

75 NER_CONCEPTS_PLUS_LLM_TYPES = "ner_concepts_plus_llm_types" 

76 LLM_TAGGED = "llm_tagged" 

77 

78 

79class TableModel(StrEnum): 

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

81 

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

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

84 """ 

85 

86 DISABLED = "disabled" 

87 TATR = "tatr" 

88 SLANET_AUTO = "slanet_auto" 

89 SLANET_PLUS = "slanet_plus" 

90 SLANET_WIRED = "slanet_wired" 

91 SLANET_WIRELESS = "slanet_wireless" 

92 

93 

94class KvCacheType(StrEnum): 

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

96 

97 F16 = "f16" 

98 F32 = "f32" 

99 Q8_0 = "q8_0" 

100 Q4_0 = "q4_0" 

101 

102 

103# Bytes per KV element for memory budgeting, from llama.cpp's block layouts: 

104# q8_0 stores 32 elements in 34 bytes (2-byte scale + 32 data bytes) and q4_0 

105# stores 32 elements in 18 bytes (2-byte scale + 16 bytes of packed nibbles). 

106# Rounding q4_0 up to a whole byte charged its cache at almost twice the real 

107# size and nearly halved the context window the dynamic picker granted. 

108KV_CACHE_TYPE_BYTES: dict[KvCacheType, float] = { 

109 KvCacheType.F16: 2.0, 

110 KvCacheType.F32: 4.0, 

111 KvCacheType.Q8_0: 34 / 32, 

112 KvCacheType.Q4_0: 18 / 32, 

113}