sprout/config/sprout.yaml at main · ghzhost/sprout · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
# Sprout configuration — the only file an adopter edits to point Sprout at a different
# corpus, switch the offline generator for Claude-on-Bedrock, or add a language.
# Every key is validated (extra keys fail loudly). Secrets come from the environment.

corpus:
  path: corpus/processed          # cleaned, chunked passages
  glob: "**/*.md"
  manifest: corpus/manifest.yaml  # per-doc source / license / fetch_date / language
  default_language: en

chunk:
  max_words: 120
  overlap_words: 24

retrieval:
  top_k: 6
  min_score: 0.12                 # below this, refuse rather than pad
  embedding_dim: 512
  embedding_provider: deterministic   # deterministic (offline hashing) | static (offline curated EN/ES vectors, ADR-0017) | bedrock (Titan)
  hybrid: true                    # fuse dense + BM25 via reciprocal rank fusion
  rrf_k: 60
  dedup_threshold: 0.92
  topic_filter: true              # scope to the named species when the question names one
  species_aliases:                # common/Spanish names -> corpus species slug
    potos: pothos
    potho: pothos
    "lengua de suegra": snake-plant
    "planta de serpiente": snake-plant
    sansevieria: snake-plant
    serpiente: snake-plant
    "lirio de la paz": peace-lily
    espatifilo: peace-lily
    filodendro: philodendron
    "ficus lira": fiddle-leaf-fig
    "arbol del caucho": rubber-plant
    "ficus elastica": rubber-plant
    dracena: dracaena
    drácena: dracaena
    hiedra: english-ivy
    "arbol de jade": jade-plant
    jade: jade-plant
    "helecho de boston": boston-fern
    helecho: boston-fern
    orquidea: orchid
    cinta: spider-plant
    malamadre: spider-plant
    calatea: calathea
    "planta zz": zz-plant
    zamioculca: zz-plant
    sabila: aloe

generation:
  provider: deterministic         # deterministic (extractive) | bedrock | anthropic
  max_sentences: 3
  relevance_floor: 0.30
  support_overlap: 0.66           # citation guard: a sentence must be this covered by its chunk
  # model: anthropic.claude-haiku-4-5-20251001-v1:0   # provider: bedrock
  # model: claude-haiku-4-5-20251001                  # provider: anthropic (native)
  region: us-west-2
  max_cost_usd: 0.05             # per-answer preflight ceiling; unpriced models refuse
  redact_query_pii: false         # redact PII before sending to a network provider
  # EXP-04 (docs/adr/0018): an additional cross-encoder NLI entailment gate behind the
  # citation guard, for the cloud path only — provider must not be "deterministic" and
  # nli.model_sha256 must be pinned, enforced at config load (fail closed).
  support_verifier: lexical       # lexical (default, zero-dependency) | nli
  # nli:
  #   model_id: cross-encoder/nli-deberta-v3-xsmall
  #   revision: main               # pin to a commit sha, not a branch, in production
  #   onnx_filename: onnx/model.onnx
  #   model_sha256: "<sha256 of the pinned onnx file — required to enable nli>"
  #   entail_threshold: 0.5
  #   entailment_label_index: 1    # this checkpoint's "entailment" index; verify per model

identification:                   # photo plant-ID -> grounded care lookup
  provider: offline               # offline (no network, always falls back) | plantnet
  min_confidence: 0.30            # below this visual-match score, ask for the name instead
  top_k: 5
  endpoint: https://my-api.plantnet.org/v2/identify/all   # the one allowlisted vision egress
  timeout_s: 30.0
  max_image_bytes: 8000000
  # The Pl@ntNet API key is read from PLANTNET_API_KEY (env only). A returned species is a
  # selector routed back through the cited corpus, never presented as a horticultural fact.

reminders:                        # local, offline watering/fertilizing reminders
  path: var/reminders.json        # stored on this machine only; nothing is uploaded
  max_reminders: 200
  default_intervals:
    water: 7
    fertilize: 30
    repot: 365
    mist: 3
    rotate: 14

confidence:
  abstain_threshold: 0.25         # below this, refuse rather than guess (per ADR-0012)
  low_confidence_threshold: 0.50  # below this, answer but flag for human review (per ADR-0012)
  reliability_bins: 10

review:                           # local review console (EXP-17): OFF by default -- see
  enabled: false                  # docs/RESPONSIBLE-TECH-AUDITS.md section C and ADR-0020 before
  path: var/review/queue.json     # enabling. Stores question text on this machine only;
  max_items: 500                  # nothing is uploaded. A maintainer opts in with `enabled:
  capture_on_low_confidence: true # true` to queue flagged/refused traces for `sprout review`.
  capture_on_refusal: true

languages:
  supported: [en, es]             # first entry is the reference language for parity

server:
  host: 127.0.0.1
  port: 8000
  max_question_chars: 500

observability:
  log_format: text                # text | json (Tier C offline CLI; Tier A for the API)
  tier: C                         # C (offline CLI, default) | A (the optional serverless
                                   # API, infra/ — turns on OTel traces/RED metrics; see
                                   # docs/ROADMAP.md's Observability tier section)
  service_name: sprout
  service_version: 0.1.0
  deployment_environment: production

store:
  path: var/index.json