Split social-read model from local reply lane
This commit is contained in:
parent
1c8303aa5a
commit
0e73ad6e18
8 changed files with 62 additions and 3 deletions
|
|
@ -31,6 +31,7 @@ SOURCE_REPO_PATTERNS=*
|
|||
SOURCE_REPO_INCLUDE_PREFIXES=AetheriaLore:Aetheria/
|
||||
LOCAL_LLM_OLLAMA_BASE_URL=http://127.0.0.1:11434
|
||||
LOCAL_LLM_OLLAMA_MODEL=qwen3.5:9b
|
||||
LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL=
|
||||
LOCAL_LLM_OLLAMA_TIMEOUT_MS=120000
|
||||
LOCAL_LLM_OLLAMA_KEEP_ALIVE=10m
|
||||
LOCAL_LLM_OLLAMA_THINK=low
|
||||
|
|
|
|||
|
|
@ -78,6 +78,7 @@ If you want the remote Ollama generation path too, configure:
|
|||
|
||||
- `LOCAL_LLM_OLLAMA_BASE_URL`
|
||||
- `LOCAL_LLM_OLLAMA_MODEL`
|
||||
- `LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL`
|
||||
- `LOCAL_LLM_OLLAMA_TIMEOUT_MS`
|
||||
- `LOCAL_LLM_OLLAMA_KEEP_ALIVE`
|
||||
- `LOCAL_LLM_OLLAMA_THINK`
|
||||
|
|
@ -86,6 +87,7 @@ If you want the remote Ollama generation path too, configure:
|
|||
|
||||
`LOCAL_LLM_ALLOW_PUBLIC=false` is the sane default. It keeps the local model restricted to admins and the configured owner while you test.
|
||||
`LOCAL_LLM_OLLAMA_THINK=low` is a good default if you want better phrasing and a little more initiative without dragging reply time into the swamp.
|
||||
`LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL` is optional. Leave it blank to reuse the main local reply model, or set it when you want the room-reading sidecar on a different model than the public/local response lane.
|
||||
|
||||
If you want a private local persona instead, point `STYLE_PACK_PATH` and `SYSTEM_MESSAGES_PATH` at your own ignored local files.
|
||||
|
||||
|
|
|
|||
|
|
@ -762,7 +762,7 @@ async function inferSituationalSocialRead(
|
|||
channelName: options.guildContext.channelName ?? null,
|
||||
promptExcerpt: truncate(options.prompt, 280),
|
||||
recentMessageCount: input.recentMessages.length,
|
||||
model: options.config.localLlm.ollamaModel,
|
||||
model: options.config.localLlm.socialReadOllamaModel,
|
||||
summary: read.summary,
|
||||
roomTone: read.roomTone,
|
||||
speakerCurrentRead: read.speakerCurrentRead,
|
||||
|
|
@ -784,7 +784,7 @@ async function inferSituationalSocialRead(
|
|||
channelId: options.guildContext.channelId,
|
||||
channelName: options.guildContext.channelName ?? null,
|
||||
promptExcerpt: truncate(options.prompt, 280),
|
||||
model: options.config.localLlm.ollamaModel,
|
||||
model: options.config.localLlm.socialReadOllamaModel,
|
||||
errorMessage:
|
||||
error instanceof Error ? error.message : "Unexpected situational social read failure.",
|
||||
},
|
||||
|
|
|
|||
|
|
@ -150,7 +150,7 @@ export async function startBot(): Promise<void> {
|
|||
const contextBuilder = new ContextBuilder();
|
||||
const situationalSocialReadInferer = new OllamaSituationalSocialReadInferer({
|
||||
ollamaBaseUrl: config.localLlm.ollamaBaseUrl,
|
||||
ollamaModel: config.localLlm.ollamaModel,
|
||||
ollamaModel: config.localLlm.socialReadOllamaModel,
|
||||
ollamaTimeoutMs: config.localLlm.ollamaTimeoutMs,
|
||||
ollamaKeepAlive: config.localLlm.ollamaKeepAlive,
|
||||
ollamaNumCtx: config.localLlm.ollamaNumCtx,
|
||||
|
|
|
|||
|
|
@ -76,6 +76,7 @@ const envSchema = z.object({
|
|||
SOURCE_REPO_INCLUDE_PREFIXES: z.string().default("AetheriaLore:Aetheria/"),
|
||||
LOCAL_LLM_OLLAMA_BASE_URL: z.string().url().default("http://127.0.0.1:11434"),
|
||||
LOCAL_LLM_OLLAMA_MODEL: z.string().min(1).default("qwen3.5:9b"),
|
||||
LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL: optionalNonEmptyString,
|
||||
LOCAL_LLM_OLLAMA_TIMEOUT_MS: z.coerce.number().int().positive().default(120000),
|
||||
LOCAL_LLM_OLLAMA_KEEP_ALIVE: z.string().min(1).default("10m"),
|
||||
LOCAL_LLM_OLLAMA_THINK: z.string().default("false"),
|
||||
|
|
@ -151,6 +152,7 @@ export interface AppConfig {
|
|||
localLlm: {
|
||||
ollamaBaseUrl: string;
|
||||
ollamaModel: string;
|
||||
socialReadOllamaModel: string;
|
||||
ollamaTimeoutMs: number;
|
||||
ollamaKeepAlive: string;
|
||||
ollamaThink: OllamaThinkMode;
|
||||
|
|
@ -326,6 +328,8 @@ export function loadConfig(env: NodeJS.ProcessEnv = process.env): AppConfig {
|
|||
localLlm: {
|
||||
ollamaBaseUrl: parsed.LOCAL_LLM_OLLAMA_BASE_URL,
|
||||
ollamaModel: parsed.LOCAL_LLM_OLLAMA_MODEL,
|
||||
socialReadOllamaModel:
|
||||
parsed.LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL ?? parsed.LOCAL_LLM_OLLAMA_MODEL,
|
||||
ollamaTimeoutMs: parsed.LOCAL_LLM_OLLAMA_TIMEOUT_MS,
|
||||
ollamaKeepAlive: parsed.LOCAL_LLM_OLLAMA_KEEP_ALIVE,
|
||||
ollamaThink: parseOllamaThinkMode(parsed.LOCAL_LLM_OLLAMA_THINK),
|
||||
|
|
|
|||
|
|
@ -822,6 +822,14 @@ try {
|
|||
}
|
||||
$localLlmUrl = if ($config.ContainsKey("LOCAL_LLM_OLLAMA_BASE_URL")) { $config["LOCAL_LLM_OLLAMA_BASE_URL"] } else { "" }
|
||||
$localLlmModel = if ($config.ContainsKey("LOCAL_LLM_OLLAMA_MODEL")) { $config["LOCAL_LLM_OLLAMA_MODEL"] } else { "" }
|
||||
$localLlmSocialReadModel = if (
|
||||
$config.ContainsKey("LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL") -and
|
||||
-not [string]::IsNullOrWhiteSpace($config["LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL"])
|
||||
) {
|
||||
$config["LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL"]
|
||||
} else {
|
||||
$localLlmModel
|
||||
}
|
||||
$localLlmEnabled = ($enabledProviders -contains "local_llm") -or (
|
||||
$config.ContainsKey("LOCAL_LLM_ALLOW_PUBLIC") -and [string]$config["LOCAL_LLM_ALLOW_PUBLIC"] -eq "true"
|
||||
)
|
||||
|
|
@ -842,16 +850,46 @@ try {
|
|||
models = $localLlmModels
|
||||
}
|
||||
}
|
||||
|
||||
if ([string]::IsNullOrWhiteSpace($localLlmSocialReadModel)) {
|
||||
Add-Check -Name "local_llm.social_read_model" -Status "warning" -Detail "Social-read sidecar model is blank, so it will fall back to the main local LLM model." -Data @{
|
||||
url = $localLlmUrl
|
||||
model = $localLlmModel
|
||||
}
|
||||
} elseif ($localLlmModels -contains $localLlmSocialReadModel) {
|
||||
$socialReadDetail = if ($localLlmSocialReadModel -eq $localLlmModel) {
|
||||
"Social-read sidecar reuses the main local LLM model $localLlmSocialReadModel."
|
||||
} else {
|
||||
"Social-read sidecar is reachable and has $localLlmSocialReadModel."
|
||||
}
|
||||
|
||||
Add-Check -Name "local_llm.social_read_model" -Status "passed" -Detail $socialReadDetail -Data @{
|
||||
url = $localLlmUrl
|
||||
model = $localLlmSocialReadModel
|
||||
}
|
||||
} else {
|
||||
Add-Check -Name "local_llm.social_read_model" -Status "failed" -Detail "Local LLM Ollama is reachable but missing social-read model $localLlmSocialReadModel." -Data @{
|
||||
url = $localLlmUrl
|
||||
models = $localLlmModels
|
||||
}
|
||||
}
|
||||
} catch {
|
||||
Add-Check -Name "local_llm.model" -Status "failed" -Detail "Local LLM Ollama check failed: $($_.Exception.Message)" -Data @{
|
||||
url = $localLlmUrl
|
||||
model = $localLlmModel
|
||||
}
|
||||
Add-Check -Name "local_llm.social_read_model" -Status "failed" -Detail "Social-read Ollama check failed: $($_.Exception.Message)" -Data @{
|
||||
url = $localLlmUrl
|
||||
model = $localLlmSocialReadModel
|
||||
}
|
||||
}
|
||||
} else {
|
||||
Add-Check -Name "local_llm.model" -Status "passed" -Detail "Local LLM check skipped because the provider is not enabled for this runtime." -Data @{
|
||||
enabledProviders = $enabledProviders
|
||||
}
|
||||
Add-Check -Name "local_llm.social_read_model" -Status "passed" -Detail "Social-read model check skipped because the local LLM provider is not enabled for this runtime." -Data @{
|
||||
enabledProviders = $enabledProviders
|
||||
}
|
||||
}
|
||||
|
||||
$sourceRepoRoot = if ($config.ContainsKey("SOURCE_REPO_ROOT")) { $config["SOURCE_REPO_ROOT"] } else { "" }
|
||||
|
|
|
|||
|
|
@ -478,6 +478,7 @@ $ragOllamaModel = if ($config.ContainsKey("RAG_OLLAMA_MODEL")) { $config["RAG_OL
|
|||
$localLlmEnabled = if ($config.ContainsKey("ENABLED_PROVIDERS")) { $config["ENABLED_PROVIDERS"] -like "*local_llm*" } else { $false }
|
||||
$localLlmBaseUrl = if ($config.ContainsKey("LOCAL_LLM_OLLAMA_BASE_URL")) { $config["LOCAL_LLM_OLLAMA_BASE_URL"] } else { "http://127.0.0.1:11434" }
|
||||
$localLlmModel = if ($config.ContainsKey("LOCAL_LLM_OLLAMA_MODEL")) { $config["LOCAL_LLM_OLLAMA_MODEL"] } else { "qwen3.5:9b" }
|
||||
$localLlmSocialReadModel = if ($config.ContainsKey("LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL") -and -not [string]::IsNullOrWhiteSpace($config["LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL"])) { $config["LOCAL_LLM_SOCIAL_READ_OLLAMA_MODEL"] } else { $localLlmModel }
|
||||
|
||||
if ($stateStorageBackend -eq "postgres") {
|
||||
$status.stage = "postgres"
|
||||
|
|
@ -514,6 +515,18 @@ if ($localLlmEnabled) {
|
|||
$status.stage = "local_llm"
|
||||
Write-StatusFile -Path $statusPath -Status $status
|
||||
$status.localLlmOllama = Ensure-OllamaEndpoint -BaseUrl $localLlmBaseUrl -Model $localLlmModel -Label "Local LLM Ollama" -LogDir $logDir
|
||||
|
||||
if ($localLlmSocialReadModel -eq $localLlmModel) {
|
||||
$status.localLlmSocialReadOllama = @{
|
||||
url = $localLlmBaseUrl
|
||||
model = $localLlmSocialReadModel
|
||||
healthy = $true
|
||||
mirrorsLocalLlm = $true
|
||||
}
|
||||
} else {
|
||||
$status.localLlmSocialReadOllama = Ensure-OllamaEndpoint -BaseUrl $localLlmBaseUrl -Model $localLlmSocialReadModel -Label "Local LLM social-read Ollama" -LogDir $logDir
|
||||
}
|
||||
|
||||
Write-StatusFile -Path $statusPath -Status $status
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -32,3 +32,4 @@
|
|||
{"ts":"2026-05-02T22:16:00+01:00","type":"architecture","status":"ok","note":"Ollama sidecar calls are more reliable under actual JSON-schema structured outputs than bare JSON mode. The situational reader now uses schema-constrained `format` payloads with deterministic temperature and can carry room-read fields plus pronoun evidence in a single no-thinking pass."}
|
||||
{"ts":"2026-05-02T21:52:56Z","type":"implementation","status":"ok","note":"Situational sidecar now emits transcript-wide Ghostlight-style participant reads with durable binary label families plus ephemeral situational state; roster-constrained schema and full configured Ollama timeout are needed to keep multi-person reads viable on a busy remote box."}
|
||||
{"ts":"2026-05-02T22:58:31Z","type":"research","status":"ok","note":"Gemma3 12B Q4_0 on the 4060 Ti box handles the transcript-wide social-read task once the schema sticks to a simpler subset; Ollama chokes on roster-constrained prefixItems/const schema but both Gemma and Qwen work with local roster reconciliation, and Gemma slightly beats Qwen on latency for the tested slice."}
|
||||
{"ts":"2026-05-02T23:40:40Z","type":"implementation","status":"ok","note":"Split room-read sidecar model selection from the main local reply model so Gemma can drive transcript profiling without changing the public/local answer lane."}
|
||||
|
|
|
|||
Loading…
Reference in a new issue