Bias local retrieval toward Discord history

This commit is contained in:
Meta Vens 2026-05-02 19:58:43 +01:00
commit 30e860c38e
3 changed files with 6 additions and 2 deletions

View file

@ -28,6 +28,8 @@ export function buildSystemPrompt(context: ContextBundle): string {
"Treat the attached interaction memory as a non-clinical behavioral read, not a diagnosis. Use the remembered dimensions, traits, and guidance to adapt tone, pacing, firmness, structure, and warmth to the person in front of you.",
"Be steady with anxious or validation-seeking speakers, grounding with grandiose ones, transparent with suspicious ones, structured with rigid or obsessive ones, and firmer with controlling, contemptuous, or boundary-pushing ones.",
"When the answer depends on archived Discord history or indexed repo/lore context, use the available read-only tools instead of guessing.",
"If the user is asking about a real incident, personal history, prior server drama, something somebody said here, or a historical event discussed in Discord, start with search_history.",
"Use search_sources for code, docs, lore, repo content, or authored project material. Do not treat it as the first stop for real-life incidents or prior server conversations.",
"If you need to target a specific indexed repo and do not know the valid repo names yet, call list_indexed_repos before search_sources.",
renderSourceGroundingInstructions(context, false),
"Do not claim to have performed searches or tool calls beyond the material actually executed in this run.",

View file

@ -22,7 +22,7 @@ export function buildToolDefinitions(): OllamaToolCallDefinition[] {
function: {
name: "search_history",
description:
"Search archived Discord history for prior conversations, decisions, preferences, and discussion fragments.",
"Search archived Discord history for prior conversations, real incidents, personal stories, decisions, preferences, relationship history, and discussion fragments from the server.",
parameters: {
type: "object",
required: ["query"],
@ -78,7 +78,7 @@ export function buildToolDefinitions(): OllamaToolCallDefinition[] {
function: {
name: "search_sources",
description:
"Search indexed source trees and lore repositories for code, documentation, and lore passages. Omit repoName to search across all indexed repos, or call list_indexed_repos first if you need valid repo names.",
"Search indexed source trees and lore repositories for code, documentation, authored worldbuilding, and project material. Use this for repo/lore questions, not as the first stop for real-life incidents or prior Discord conversations. Omit repoName to search across all indexed repos, or call list_indexed_repos first if you need valid repo names.",
parameters: {
type: "object",
required: ["query"],

View file

@ -20,3 +20,5 @@
{"ts":"2026-05-02T17:34:02Z","type":"implementation","status":"ok","note":"Layered private system message catalogs over the tracked base, deferred slow slash replies before provider execution, and chunked Discord output so missing keys and overlong local_llm replies stop failing live interactions."}
{"ts":"2026-05-02T17:39:46Z","type":"implementation","status":"ok","note":"Source-grounding alias matching now ignores overly short repo components so social prompts mentioning Void stop tripping fake repo-evidence requirements, and direct messages intentionally count as prompts without requiring a mention."}
{"ts":"2026-05-02T18:28:04Z","type":"implementation","status":"ok","note":"Pinned the owner Codex lane to a configurable CODEX_MODEL defaulting to gpt-5.4 because the local CLI still faceplants on gpt-5.5, and relaxed source-grounding from a lexical hard gate into an advisory hint so the model can decide when repo/lore tool use is actually needed."}
{"ts":"2026-05-02T18:41:02Z","type":"bugfix","status":"ok","note":"Direct local_llm replies could emit raw JSON/tool payload fragments as final Discord answers; added structured-output rewrite guard and audit output previews for direct provider responses."}
{"ts":"2026-05-02T18:58:38Z","type":"bugfix","status":"ok","note":"Local LLM could choose search_sources for real-life incident questions and get Ghostlight narrative hits; prompt/tool guidance now biases prior server drama and personal-history questions toward search_history before repo/lore search."}