4da32ad102
Local Ollama has three embedding models loaded: nomic-embed-text:latest 137M 768d (previous default) nomic-embed-text-v2-moe:latest 475M 768d (this commit's default) qwen3-embedding:latest 7.6B 4096d (would require dim change) v2-moe is a drop-in upgrade — same 768 dim, 3.5× more params, MoE architecture. Workers index doesn't need rebuilding, just future ingests embed with the stronger model. Run #005 result on the multi-coord stress suite: Diversity (same-role-across-contracts): 0.080 → 0.000 (n=9) → MoE is more discriminating: zero worker overlap across Milwaukee / Indianapolis / Chicago for shared role names. The geo + cert + skill context fully separates worker pools. Different-roles-same-contract: 0.013 → 0.036 (still ~96% diff) Determinism: 1.000 (unchanged) Verbatim handover: 4/4 (100%) Paraphrase handover: 4/4 (100%) 200-worker swap: Jaccard 0.000 (unchanged — still perfect) Fresh-resume verify: STILL doesn't surface fresh workers in top-8. With v2-moe, distances increased (top-1 = 0.43–0.65 vs v1's 0.25–0.39) — the embedder is MORE discriminating, but the fresh worker's vector still doesn't outrank the 8th-best existing worker. Now suspect of being an HNSW post-build add issue (coder/hnsw incremental adds can land in hard-to-reach graph regions, not an embedder problem). Better embedder didn't fix it; needs a different strategy: full index rebuild after fresh adds, or explicit playbook-layer score boost for fresh workers, or hybrid (keyword + semantic) retrieval. Phase 3 investigation. Cost: ingest is ~5× slower (workers 20s→100s; ethereal 35s→112s). Acceptable for the quality jump on diversity. Real production with incremental ingest won't pay this once-per-deploy. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
156 lines
6.0 KiB
TOML
156 lines
6.0 KiB
TOML
# Lakehouse-Go config — G0 dev defaults. Overrides via env are a
|
||
# G1+ concern; for G0 edit this file and restart the affected service.
|
||
|
||
# G0 dev ports — shifted to 3110+ so the Go services run alongside
|
||
# the live Rust lakehouse on 3100/3201-3204 without colliding. G5
|
||
# (demo cutover) flips gateway back to 3100 when Rust retires.
|
||
[gateway]
|
||
bind = "127.0.0.1:3110"
|
||
storaged_url = "http://127.0.0.1:3211"
|
||
catalogd_url = "http://127.0.0.1:3212"
|
||
ingestd_url = "http://127.0.0.1:3213"
|
||
queryd_url = "http://127.0.0.1:3214"
|
||
vectord_url = "http://127.0.0.1:3215"
|
||
embedd_url = "http://127.0.0.1:3216"
|
||
pathwayd_url = "http://127.0.0.1:3217"
|
||
matrixd_url = "http://127.0.0.1:3218"
|
||
observerd_url = "http://127.0.0.1:3219"
|
||
chatd_url = "http://127.0.0.1:3220"
|
||
|
||
[storaged]
|
||
bind = "127.0.0.1:3211"
|
||
|
||
[catalogd]
|
||
bind = "127.0.0.1:3212"
|
||
storaged_url = "http://127.0.0.1:3211"
|
||
|
||
[ingestd]
|
||
bind = "127.0.0.1:3213"
|
||
storaged_url = "http://127.0.0.1:3211"
|
||
catalogd_url = "http://127.0.0.1:3212"
|
||
# CSV uploads are ~4-6× the resulting Parquet. 256 MiB cap keeps the in-memory
|
||
# parse + Arrow + Parquet output footprint bounded. Bump for known large
|
||
# datasets (e.g. workers_500k → 344 MiB CSV needs 512 MiB).
|
||
max_ingest_bytes = 268435456
|
||
|
||
[vectord]
|
||
bind = "127.0.0.1:3215"
|
||
# Optional — set to empty string to disable persistence (dev/test).
|
||
storaged_url = "http://127.0.0.1:3211"
|
||
|
||
[embedd]
|
||
bind = "127.0.0.1:3216"
|
||
# G2: Ollama local. G3+ may swap in OpenAI/Voyage by changing
|
||
# this URL + the wire format inside the provider.
|
||
provider_url = "http://localhost:11434"
|
||
default_model = "nomic-embed-text-v2-moe"
|
||
|
||
[queryd]
|
||
bind = "127.0.0.1:3214"
|
||
catalogd_url = "http://127.0.0.1:3212"
|
||
secrets_path = "/etc/lakehouse/secrets-go.toml"
|
||
refresh_every = "30s"
|
||
|
||
[pathwayd]
|
||
bind = "127.0.0.1:3217"
|
||
# Empty = in-memory only (dev/test). Production sets a path under
|
||
# /var/lib/lakehouse/pathway/state.jsonl so traces survive restart.
|
||
persist_path = ""
|
||
|
||
[matrixd]
|
||
bind = "127.0.0.1:3218"
|
||
# matrixd calls embedd (query-text → vector) and vectord (per-corpus
|
||
# search) directly. Localhost defaults; in distributed deployments
|
||
# these point at the gateway's upstream addresses.
|
||
embedd_url = "http://127.0.0.1:3216"
|
||
vectord_url = "http://127.0.0.1:3215"
|
||
|
||
[observerd]
|
||
bind = "127.0.0.1:3219"
|
||
# Empty = in-memory only (dev/test). Production sets a path under
|
||
# /var/lib/lakehouse/observer/ops.jsonl so ops survive restart.
|
||
persist_path = ""
|
||
|
||
[chatd]
|
||
# LLM chat dispatcher (Phase 4). Routes /v1/chat to the right provider
|
||
# based on model name prefix or :cloud suffix:
|
||
# ollama/<model> → local Ollama (no auth)
|
||
# ollama_cloud/<model> → Ollama Cloud
|
||
# <model>:cloud → Ollama Cloud (suffix variant)
|
||
# openrouter/<vendor>/<m> → OpenRouter
|
||
# opencode/<model> → OpenCode (Zen+Go unified)
|
||
# kimi/<model> → Kimi For Coding
|
||
# bare names (e.g. qwen3.5:latest) → local Ollama (default)
|
||
bind = "127.0.0.1:3220"
|
||
ollama_url = "http://localhost:11434"
|
||
|
||
# Per-provider key resolution: env var first, then .env file fallback.
|
||
# Empty file path skips the file lookup. systemd EnvironmentFile=
|
||
# loads these natively, so service runtime sees the env vars.
|
||
ollama_cloud_key_env = "OLLAMA_CLOUD_KEY"
|
||
openrouter_key_env = "OPENROUTER_API_KEY"
|
||
opencode_key_env = "OPENCODE_API_KEY"
|
||
kimi_key_env = "KIMI_API_KEY"
|
||
|
||
ollama_cloud_key_file = "/etc/lakehouse/ollama_cloud.env"
|
||
openrouter_key_file = "/etc/lakehouse/openrouter.env"
|
||
opencode_key_file = "/etc/lakehouse/opencode.env"
|
||
kimi_key_file = "/etc/lakehouse/kimi.env"
|
||
|
||
# Per-call timeout (seconds). Cloud reasoning models can take >60s
|
||
# for long prompts, so 180 is the default.
|
||
timeout_secs = 180
|
||
|
||
[s3]
|
||
endpoint = "http://localhost:9000"
|
||
region = "us-east-1"
|
||
bucket = "lakehouse-go-primary" # G0 dedicated bucket so Rust + Go coexist
|
||
access_key_id = "" # populated by SecretsProvider from /etc/lakehouse/secrets-go.toml
|
||
secret_access_key = "" # ditto
|
||
use_path_style = true
|
||
|
||
[log]
|
||
level = "info"
|
||
|
||
# Model tier registry — names map to actual model IDs per the small-
|
||
# model pipeline architecture (project_small_model_pipeline_vision.md).
|
||
# Bumping a model means editing one line here, not hunting through code.
|
||
#
|
||
# Tier philosophy:
|
||
# - local_* : on-box Ollama. Inner-loop hot path. Repeated calls.
|
||
# - cloud_* : Ollama Cloud (Pro plan). Larger context, fail-up tier.
|
||
# - frontier_* : OpenRouter / OpenCode. Rate-limited, billed per call.
|
||
# Reserved for blockers and full-scope reviews.
|
||
#
|
||
# weak_models is the codified "local-hot-path eligible" list that the
|
||
# matrix.downgrade gate reads. A model in this list bypasses the
|
||
# strong-model downgrade rule (it's already weak — no need to downgrade
|
||
# corpora further).
|
||
[models]
|
||
# Tier 1 — local hot path
|
||
local_fast = "qwen3.5:latest"
|
||
local_embed = "nomic-embed-text-v2-moe" # 475M MoE, drop-in upgrade from 137M v1 — verified 2026-04-30 same 768-dim
|
||
# local_judge stays on qwen2.5:latest — qwen3.5:latest is a vision-SSM
|
||
# build with 256K context that runs ~30s per judge call against the
|
||
# playbook_lift loop (verified 2026-04-30). qwen2.5:latest at ~1s/call
|
||
# is 30× faster and held lift theory across the 21-query reality test
|
||
# (7/8 lift, 87.5%). The 8de94eb "bump qwen2.5 → qwen3.5" was a casual
|
||
# version-up; this revert is workload-specific.
|
||
local_judge = "qwen2.5:latest"
|
||
local_review = "qwen3.5:latest"
|
||
|
||
# Tier 2 — Ollama Cloud (Pro). kimi-k2:1t still upstream-broken;
|
||
# deepseek/kimi-k2.6/qwen3-coder are the working primaries.
|
||
cloud_judge = "kimi-k2.6:cloud"
|
||
cloud_review = "qwen3-coder:480b"
|
||
cloud_strong = "deepseek-v3.2"
|
||
|
||
# Tier 3 — frontier. Use sparingly; rate-limited + per-call billing.
|
||
frontier_review = "openrouter/anthropic/claude-opus-4-7"
|
||
frontier_arch = "openrouter/moonshotai/kimi-k2-0905"
|
||
frontier_strong = "openrouter/openai/gpt-5"
|
||
frontier_free = "opencode/claude-opus-4-7"
|
||
|
||
# Local-hot-path eligible — matrix.downgrade bypass list.
|
||
weak_models = ["qwen3.5:latest", "qwen3:latest"]
|