Skip to content

config.md

Latest commit

 

History

History
276 lines (212 loc) · 11 KB

config.md

File metadata and controls

276 lines (212 loc) · 11 KB

Configuration Reference

Chronicles is configured via chronicles/config.yaml. All sections are optional — sensible defaults are used when absent.

Full Example

# ──────────────────────────────────────────────────
# LLM Provider
# ──────────────────────────────────────────────────
llm:
  provider: claude-code          # claude-code | copilot-cli | ollama
  model: claude-haiku-4-5-20251001
  max_concurrent: 8              # parallel LLM calls (extraction, enrichment)

  # Per-step overrides (optional).
  # Each step inherits from the global llm settings above.
  # Only specify the fields you want to override.
  extract:
    provider: ollama
    model: gemma4:26b
  enrich:
    model: claude-haiku-4-5-20251001
  signals:
    provider: ollama
    model: qwen3.5:9b
  similarity:
    model: claude-haiku-4-5-20251001

# ──────────────────────────────────────────────────
# Ollama Settings (only used when provider is "ollama")
# ──────────────────────────────────────────────────
ollama:
  base_url: http://localhost:11434   # Ollama server address
  timeout: 300                       # seconds per request
  temperature: 0.0                   # 0 = greedy decoding
  num_ctx: 0                         # context window size (0 = model default)
  num_predict: 0                     # max output tokens (0 = model default)

# ──────────────────────────────────────────────────
# Transcript Sources
# ──────────────────────────────────────────────────
sources:
  - claude-code
  - copilot-cli
  - copilot-vscode

# ──────────────────────────────────────────────────
# Wiki Confidence Lifecycle
# ──────────────────────────────────────────────────
confidence:
  promotion_threshold: 3       # sessions confirming an article before promotion to high

# ──────────────────────────────────────────────────
# Confidence Decay
# ──────────────────────────────────────────────────
decay:
  high_to_medium_days: 180     # days without confirmation before high → medium
  medium_to_low_days: 270      # days without confirmation before medium → low
  archive_after_days: 365      # days without confirmation before archiving

# ──────────────────────────────────────────────────
# Record Archiving
# ──────────────────────────────────────────────────
archive:
  after_days: 90               # move records to archives/ after N days

# ──────────────────────────────────────────────────
# Enrichment
# ──────────────────────────────────────────────────
enrich:
  enabled: true                # auto-enrich after ingest (category summaries)

# ──────────────────────────────────────────────────
# Similarity / Deduplication
# ──────────────────────────────────────────────────
similarity:
  engine: tfidf                # tfidf | llm
  threshold: 0.2               # minimum score to flag as potential duplicate
  confirm_engine: ""           # optional second-stage engine (e.g. "llm")
  confirm_threshold: 0.7       # threshold for second-stage confirmation

# ──────────────────────────────────────────────────
# Knowledge Gaps Detection
# ──────────────────────────────────────────────────
gaps:
  enabled: true
  git_lookback_days: 90        # how far back to scan git history

# ──────────────────────────────────────────────────
# Operational Signals
# ──────────────────────────────────────────────────
signals:
  max_active: 50               # max active signals before oldest are demoted
  demoted_retention_days: 90   # days to keep demoted signals before removal
  subagents: true              # also extract signals from subagent sessions

Section Details

llm

The global LLM configuration used across all pipeline steps.

Field Type Default Description
provider string copilot-cli LLM backend: claude-code, copilot-cli, or ollama
model string gpt-5-mini Model name passed to the provider
max_concurrent int 3 Max parallel LLM calls

Per-step Overrides

Each pipeline step that uses an LLM can override provider and/or model. Unspecified fields inherit from the global llm section. max_concurrent is always inherited.

Step Used by Description
extract chronicles ingest Knowledge extraction from transcripts
enrich chronicles enrich Category summary generation
signals chronicles signals Operational signal detection
similarity chronicles lint LLM-backed duplicate detection (when similarity.engine: llm)

ollama

Configuration for the Ollama provider. Only relevant when at least one step uses provider: ollama. These settings apply globally to all ollama calls.

Field Type Default Description
base_url string http://localhost:11434 Ollama server URL. Change for remote servers or custom ports.
timeout int 300 Request timeout in seconds. Increase for large models or slow hardware.
temperature float 0.0 Sampling temperature. 0 = deterministic (greedy).
num_ctx int 0 Context window size. 0 = use model default. Set explicitly for models that support large contexts.
num_predict int 0 Max output tokens. 0 = use model default.

sources

List of transcript sources to scan when using --since. Available sources:

  • claude-code — Claude Code CLI sessions (~/.claude/projects/)
  • copilot-cli — GitHub Copilot CLI sessions
  • copilot-vscode — Copilot Chat in VS Code

confidence

Controls how wiki articles are promoted from medium to high confidence.

Field Type Default Description
promotion_threshold int 3 Number of independent sessions confirming an article before it reaches high confidence

decay

Controls automatic confidence demotion when articles are not reconfirmed.

Field Type Default Description
high_to_medium_days int 180 Days without confirmation before high → medium
medium_to_low_days int 270 Days without confirmation before medium → low
archive_after_days int 365 Days without confirmation before article is archived

archive

Field Type Default Description
after_days int 90 Move ingestion records from records/ to archives/ after this many days

enrich

Field Type Default Description
enabled bool true Automatically run enrichment after ingestion

similarity

Controls wiki deduplication during lint.

Field Type Default Description
engine string tfidf Similarity engine: tfidf (fast, local) or llm (uses configured LLM)
threshold float 0.2 Minimum similarity score to flag as potential duplicate
confirm_engine string "" Optional second-stage engine for higher-confidence dedup
confirm_threshold float 0.7 Threshold for second-stage confirmation

gaps

Field Type Default Description
enabled bool true Enable knowledge gap detection based on git history
git_lookback_days int 90 How far back to scan git history for uncovered areas

signals

Field Type Default Description
max_active int 50 Maximum active signals. Oldest are demoted when exceeded.
demoted_retention_days int 90 Days to keep demoted signals in SIGNALS.md
subagents bool true Also process subagent transcripts for signals

Use Case Configurations

Cloud API only (default)

Use Claude or Copilot as the sole provider. Simple setup, nothing to run locally.

llm:
  provider: claude-code
  model: claude-haiku-4-5-20251001
  max_concurrent: 8

Fully local with Ollama

Run everything on your own hardware. No API keys needed.

llm:
  provider: ollama
  model: gemma4:26b
  max_concurrent: 4

ollama:
  timeout: 600        # local models are slower

Hybrid: smart model for extraction, fast model for the rest

Use a capable model for the hard task (knowledge extraction) and a cheaper/faster model for simpler tasks.

llm:
  provider: claude-code
  model: claude-haiku-4-5-20251001
  max_concurrent: 8
  extract:
    model: claude-sonnet-4-5-20250514

Hybrid: local extraction, cloud enrichment

Extract locally to avoid sending full transcripts to the cloud, but use a cloud model for the smaller enrichment prompts.

llm:
  provider: claude-code
  model: claude-haiku-4-5-20251001
  max_concurrent: 8
  extract:
    provider: ollama
    model: gemma4:26b
  signals:
    provider: ollama
    model: gemma4:26b

ollama:
  timeout: 600

Remote Ollama server

Run Ollama on a beefy machine and connect from your laptop.

llm:
  provider: ollama
  model: qwen3.5:35b-a3b
  max_concurrent: 2

ollama:
  base_url: http://192.168.1.50:11434
  timeout: 600
  num_ctx: 32768