LLM-assisted KB miner: bounded candidate learning via llm.starintel.actor 27B #93

New issue

Open

opened 2026-09-20 01:19:24 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-20 01:19:24 +00:00

Owner

Parent: corpus workbench epic.

Goal

Mine the durable llm-log expert KB and corpus analytics for candidate reusable facts/rules using the operator's 27B model served by llm.starintel.actor.

Model configuration

OpenAI-compatible base URL: configurable, default https://llm.starintel.actor.
Do not guess a model ID in source.
Support an operator override plus runtime auto:27b discovery from /v1/models.
Auto discovery must select exactly one explicitly reported 27B model or fail clearly.
Credentials remain runtime secrets, never Nix-store values/capture logs.

Bounded mining loop

execute a declared indexed Tek9 query with finite page/batch size;
construct a prompt from only that bounded evidence set;
request candidate fact/rule/classification improvements from the model;
validate against typed schemas + Prolog syntax/declared predicate allowlist;
persist candidate + source IDs + model ID + prompt hash + query + timestamp;
test candidates against held-out/historical evidence;
publish/promote only through an explicit versioned rule transition.

The miner never receives the full corpus/KB.

Outputs

candidate facts
candidate Prolog rules
taxonomy/classification suggestions
failure/retry patterns
cost/quality correlations
evidence-backed report text

Acceptance

Every learned candidate is reproducible from a finite evidence manifest and can be rejected/superseded without changing raw evidence.

Parent: corpus workbench epic. ## Goal Mine the durable llm-log expert KB and corpus analytics for candidate reusable facts/rules using the operator's 27B model served by `llm.starintel.actor`. ## Model configuration - OpenAI-compatible base URL: configurable, default `https://llm.starintel.actor`. - Do not guess a model ID in source. - Support an operator override plus runtime `auto:27b` discovery from `/v1/models`. - Auto discovery must select exactly one explicitly reported 27B model or fail clearly. - Credentials remain runtime secrets, never Nix-store values/capture logs. ## Bounded mining loop 1. execute a declared indexed Tek9 query with finite page/batch size; 2. construct a prompt from only that bounded evidence set; 3. request candidate fact/rule/classification improvements from the model; 4. validate against typed schemas + Prolog syntax/declared predicate allowlist; 5. persist candidate + source IDs + model ID + prompt hash + query + timestamp; 6. test candidates against held-out/historical evidence; 7. publish/promote only through an explicit versioned rule transition. The miner never receives the full corpus/KB. ## Outputs - candidate facts - candidate Prolog rules - taxonomy/classification suggestions - failure/retry patterns - cost/quality correlations - evidence-backed report text ## Acceptance Every learned candidate is reproducible from a finite evidence manifest and can be rejected/superseded without changing raw evidence.