KrisHD_AI/README.md

Elsewhere

Elsewhere is an Express + SQLite app for account-based chat, persona management, roleplay chat sessions, and a simple persona market backed by a remote model API.

A space for every conversation. Six appearance styles (Grove, Parchment, Aurora, Current, Iris, and Graphite) each support light and dark modes. The Elsewhere identity stays consistent across styles. Internal theme and storage keys remain stable to preserve saved preferences.

Setup

Node.js 22 or newer is required.

1. Install dependencies with npm install.
2. Copy .env.example to .env and set at least SESSION_SECRET.
3. Start the app with npm run dev for local development or npm start for a normal server run.

Scripts

• npm start: run the production server entrypoint.
• npm run dev: run the server with Node watch mode.
• npm run lint: syntax-check the backend and browser modules.
• npm test: run API tests with Node's built-in test runner.

Environment Variables

• PORT: HTTP port for the Express server.
• DB_PATH: SQLite database path. Use :memory: for isolated tests.
• SESSION_SECRET: session signing secret. Required outside local throwaway environments.
• COOKIE_SECURE: set to true when serving over HTTPS. Required in production unless ALLOW_INSECURE_COOKIES=true is set.
• ALLOW_INSECURE_COOKIES: explicit production override for non-HTTPS deployments.
• TRUST_PROXY: Express trust proxy setting for reverse proxy deployments.
• SESSION_MAX_AGE_MS: cookie lifetime.
• EVENT_LOOP_LAG_*: enable and tune event-loop lag warnings.
• SLOW_REQUEST_*: enable and tune slow request warnings.
• CHAT_LIMIT: per-account maximum chat sessions.
• CHAT_HISTORY_LIMIT: number of recent messages included in model prompts.
• MODEL_API_BASE_URL: base URL of the remote model backend. The server expects /chat and /tags.
• MODEL_REQUEST_TIMEOUT_MS: legacy fallback for the inactivity timeout (default 180000 ms).
• MODEL_CONNECT_TIMEOUT_MS: time to receive Ollama response headers (default 30000 ms).
• MODEL_INACTIVITY_TIMEOUT_MS: maximum silence between received stream chunks (default 180000 ms).
• MODEL_MAX_GENERATION_MS: overall generation safety limit (default 600000 ms).
• OLLAMA_KEEP_ALIVE: Ollama model residency after a request (default 30m).
• MODEL_LIST_TIMEOUT_MS: timeout for model list requests.
• MODEL_UNLOAD_AFTER_MS: delay before idle local request bookkeeping is cleared. Model unloading is managed by Ollama through OLLAMA_KEEP_ALIVE.
• MODELS_CACHE_TTL_MS: model list cache TTL.
• AUTH_RATE_LIMIT_*: rate-limit window and thresholds for registration/login.
• CHAT_RATE_LIMIT_*: rate-limit window and thresholds for chat and retry endpoints.
• RATE_LIMIT_CLEANUP_INTERVAL_MS: interval for pruning stale in-memory rate-limit buckets.
• DROP_LEGACY_CHATS: when true, migrate any legacy chats table into chat_sessions / chat_messages and remove the old table.

Image attachments and vision

The composer supports PNG, JPEG and WebP through the image button, dropping files onto the composer, or pasting a screenshot. Text paste is unchanged. Previews show upload status and support retry/removal; sending waits until every remaining image is ready. Images also render in saved messages and open at full size in a protected browser tab.

Production configuration (the existing /app/data volume remains unchanged):

UPLOAD_DIR=/app/data/uploads
UPLOAD_MAX_SIZE_MB=10
UPLOAD_MAX_FILES=5
UPLOAD_MAX_PIXELS=30000000
VISION_MODEL=gemma3:4b

Outside production, the default directory is data/uploads; UPLOAD_DIR overrides it. Startup creates the directory and checks write permission, failing clearly rather than falling back to another location. Files use YYYY/MM/DD/<uuid>.<format> keys. Do not publish this directory as a static directory. Back up both the SQLite database and uploads directory. JSON workspace exports contain attachment references only; imports reconnect existing images owned by the same account, not portable image bytes.

Authenticated API endpoints:

• GET /uploads/config: safe client limits, without storage paths.
• POST /uploads: multipart field image, one file per request; returns {id,type:"image",originalName,mimeType,sizeBytes,width,height}.
• GET /attachments/:id: ownership-checked image bytes with private/no-store caching and nosniff; attached images also require access to a referencing conversation.
• DELETE /attachments/:id: removes an owned temporary upload only.
• /chat and /chat/stream: optional attachments: [id, ...]; image-only messages receive a default description prompt. Existing text-only payloads remain valid.

Multer bounds multipart input in memory; Sharp detects the actual format, checks the pixel limit and fully decodes it. Browser MIME types and filenames are not trusted. SVG, GIF, animated/multipage images, empty, malformed, oversized and fake images are rejected. File counts are checked again during atomic message association. Concurrent uploads per account are capped at UPLOAD_MAX_FILES, and pending uploads at four times that limit. Neither arbitrary paths nor remote image URLs are accepted.

The additive, idempotent SQLite migration creates attachments, message_attachments and upload_deletions, plus the question-specific attachment_vision_analyses cache; existing messages and files are not rewritten. Attachment rows contain owner, original filename, relative storage key, MIME/size/dimensions, temporary/attached status, creation time and optional vision model/analysis/time. Shared references preserve images across branches, edits and same-instance imports. Deleting the last referencing message/conversation (or the account) queues the file for deletion. Cleanup drains this durable queue at startup, after modifying requests, before uploads and hourly; failed file deletion is retried. Temporary uploads expire after 24 hours. Attached images do not expire. No binary or Base64 data is stored in SQLite.

Generation order is attachment ownership validation → atomic user message/reference persistence → Analyzing images… → existing Web Search decision/retrieval → final selected chat model streaming. Gemma receives each image separately as a server-side Base64 messages[].images value, together with the user's question. The existing Ollama client sends model: VISION_MODEL, think: false, a JSON Schema in format, keep-alive and the existing cancellation/timeouts. Its stream is consumed privately; only the final selected chat model's answer and opted-in reasoning reach the user.

The bounded, validated vision record is:

{
  "summary": "Concise factual overview",
  "answer": "Direct visual conclusion relevant to the user's question",
  "confidence": 0.9,
  "evidence": ["Visible details supporting the conclusion"],
  "visibleText": ["Clearly readable text"],
  "observations": [],
  "alternatives": [],
  "uncertainties": []
}

Gemma adapts its emphasis to identification, UI debugging, OCR, counting, comparison, or scene understanding. It should give the most likely supported answer, qualified when appropriate, rather than refusing because absolute certainty is impossible. answer can be null for genuinely insufficient evidence. confidence is a number from 0 to 1 (or null), not a calibrated probability. Alternatives require actual support; empty arrays are preferred to speculative lists or boilerplate limitations. Each field has a length/item limit, with at most eight evidence items and three alternatives. The helper retains its 1,600-token output budget.

Malformed JSON falls back to a bounded summary plus rawAnalysis (up to 6,000 characters), with no invented answer or confidence. Valid JSON is parsed before individual fields are bounded, so truncation cannot corrupt its structure. An empty response or unavailable Gemma fails the turn explicitly, without asking the chat model to invent image details. The cache key is attachment, configured vision model, VISION_PROMPT_VERSION (currently 2), and trimmed user question; case and internal whitespace are preserved. Records include the analysis and timestamp. Only an exact matching key is reused. Legacy model-only snapshots are retained for compatibility but never used as cache hits. Bump the prompt version when changing the analysis contract. Cache records cascade away when the attachment is deleted.

Recent-history image context remains limited to the configured image count. A new question, including a follow-up without another upload, analyzes retained images for that question. Repeated questions and regeneration reuse matching results. Different question results coexist, including across shared branches. A new question can therefore require extra Gemma calls, bounded by the existing image count.

The final model receives intact structured evidence in a separate untrusted user context message, not a truncated JSON string in a system prompt. Trusted guidance tells it to answer directly using the answer, confidence and evidence, retain real ambiguity, avoid inventing alternatives, and distinguish visible UI state from unseen backend causes. It should not mention internal tools/models unless explicitly asked. OCR and raw fallback text are data, never instructions. Analysis remains separate from visible messages; the final model never receives image bytes or filesystem paths. These are model instructions, not a deterministic guarantee of identification accuracy or final-answer wording. Automated tests cover the contract and data flow; semantic quality requires evaluation with real images and the deployed models.

Web Search and Reasoning remain independent. For requested searches, concise visual evidence helps form the query without overriding the user's search decision. Search finishes before final generation; source ordering/rendering is unchanged. The final request still uses stream: true and the existing capability-checked reasoning toggle. Stop/disconnect cancels the current vision or final generation request. Logging includes IDs, sizes, types and cache events only, never images/Base64/private analysis.

Dependencies added: multer (bounded multipart handling), sharp (content detection, dimensions and full image decoding). Install through the normal npm install/npm ci workflow, including Sharp's platform-specific optional dependencies in Docker. The application does not pull Ollama models; provision gemma3:4b separately.

Model Backend Expectations

The server is configured around an Ollama-compatible API surface:

• POST /chat: accepts { model, messages, stream: true, think: boolean, keep_alive: "30m" } and returns newline-delimited JSON chunks with optional message.content / message.thinking, ending in done: true.
• GET /tags: returns the available model list.

qwen3:14b is the default model for chat and roleplay. The model selector loads available models from /tags and selects qwen3:14b when available (otherwise the first returned model). Existing chats retain their saved model selection.

The composer has independent Web Search and Reasoning toggles, both off by default and reset on server acceptance. It sends {message, chatId, model, webSearch, reasoning} using streamed fetch to POST /chat/stream. Existing SSE events carry status (accepted/searching/generating), thinking (chunk/fullThinking), chunk (chunk/fullReply), done (reply/sources), and error. Final answers stream incrementally, including persona responses. Persona validation still runs; a corrective pass resets provisional output before streaming its replacement. Only the validated complete answer is persisted.

Every Ollama request explicitly sets stream: true, think, and keep_alive. Reasoning is requested and forwarded only when reasoning === true and the model supports the boolean thinking API. The shared conservative capability check supports Qwen3, DeepSeek R1 and DeepSeek v3.1 model IDs; unknown models, Qwen coder variants and models requiring thinking levels have the toggle disabled. Internal calls, roleplay openers and regeneration default to think: false. There is no extra model call to obtain reasoning.

Reasoning appears only after actual thinking text arrives, expanded during generation and collapsed when the answer begins. Users can reopen it; stored messages start collapsed. The optional nullable chat_messages.thinking field remains separate from answer content and sources, is excluded from future prompts and logs, and survives reloads, exports, imports and branches. Old messages remain valid.

Ollama reads reset the inactivity deadline on each received chunk; the overall safety limit still applies to active streams. Client disconnects and the existing Stop action abort upstream generation. Missing done: true, invalid streams, timeouts and upstream failures use the existing error path and never save a partial assistant response as complete. Browser EOF without a terminal SSE event is likewise an interruption. The existing UI discards temporary output on failure.

The refactor adds explicit upstream timeout handling, cached model listing, and GET /health for degraded-state visibility.

Schema Notes

Primary runtime tables are:

• users
• chat_sessions
• chat_messages
• personas
• user_settings
• persona_market

Legacy chats migration is handled during startup. New code only reads the session/message schema.

Persona And Market Flow

• Assistant and user personas are stored separately through persona_type.
• Roleplay chats can bind an assistant persona, an optional user persona, and scenario metadata to a chat session.
• Publishing copies a local persona into persona_market.
• Collecting clones a market persona back into the current user's personas table with source metadata.

Frontend Structure

The browser UI remains framework-free and uses native ES modules. Shared behavior lives under public/app/:

• api.js: JSON and streaming request helpers
• constants.js: theme and onboarding metadata
• themeController.js: validated theme/mode persistence and application
• theme-bootstrap.js: applies stored appearance before page initialization
• personaForm.js: shared persona form serialization
• statusNotice.js: shared status and error presentation
• dom.js: cached chat-page DOM references

public/style.css is a self-contained design system rebuilt around semantic tokens, shared components, and deliberate responsive breakpoints. It contains the six visual identities and their dark variants without legacy imports or precedence layers. public/legal.css contains only the legal-page layout.

Pure chat message, retry, memory, summary, and persona-state formatting lives in src/services/chatMessageHelpers.js; chatService.js remains responsible for orchestration and persistence.

Optional Web Search

The globe button next to the composer enables search for one message. It defaults to off and resets once the server accepts the message. The browser sends webSearch: true (or false) through the existing authenticated POST /chat/stream; POST /chat supports the same optional boolean. Omitting it preserves the normal chat flow. No generic tool/agent system is needed.

Search uses shouldSearch = webSearchEnabled && (explicitSearchIntent || automaticSearchDecision). Deterministic commands such as “search the internet”, “look this up online”, and “find sources” override smart skipping. Arithmetic, rewriting, explanations and questions answered by the server clock can otherwise skip search. Query cleanup removes command wording while preserving dates and years; “average lifespan” is normalized to “average life expectancy”.

The chat service retrieves up to 15 candidates per query and selects 1–5 useful sources dynamically, rejecting empty evidence and wrong-date results. SearXNG content maps to snippet (an existing non-empty snippet is also accepted). Exact-date event searches use the general index to include archives and try at most one alternate query if the first returns no usable evidence; candidates are merged and deduplicated. Current-news requests retain the news category. The same numbered selected titles, URLs and snippets go immediately before the current question in Ollama's context, into the response, and into storage. Retrieval finishes before either reasoning or answer streaming starts. Instructions require citations, supported partial answers rather than blanket claims of missing details, and honest disclosure when evidence is insufficient. Result content remains untrusted data; no result pages are fetched.

Set environment variables in the server/container environment, or load the example explicitly with node --env-file=.env server.js (the existing npm start command does not load .env automatically):

SEARXNG_URL=http://searxng:8080
WEB_SEARCH_TIMEOUT_MS=8000

SEARXNG_URL is optional and server-only. It must be an HTTP(S) base URL reachable from Express, for example over the same Docker network. Use the final URL: redirects are rejected. Missing/invalid configuration leaves ordinary chat available and returns Web search is currently unavailable. for search requests. WEB_SEARCH_TIMEOUT_MS limits the whole search response, including body download (default 8000 ms). Responses are limited to 1 MiB; each source has at most a 300-character title, 2048-character URL and 1500-character snippet. Only HTTP(S) source links without embedded credentials are accepted. No user input can override the provider URL.

Enable JSON in SearXNG's settings.yml:

search:
  formats:
    - html
    - json

The provider calls GET {SEARXNG_URL}/search?q=...&format=json using encoded query parameters. SearXNG must allow requests from the application server; disabled JSON output returns HTTP 403. See the official Search API documentation. Only the current query is sent to SearXNG, not conversation history or personas; configured SearXNG engines can forward that query to their search services.

The server-side WebSearchProvider JSDoc contract is search(query, {signal, limit}?) -> Promise<SearchResult[]>; SearXNGSearchProvider is the initial implementation. Results have {title, url, snippet}. This repository uses JavaScript ES modules, so types are documented with JSDoc rather than introducing a TypeScript build.

Search emits {status: "searching"} only when retrieval actually runs, then {status: "generating", searchState, sourceCount?} before model generation. The UI initially says “Preparing response…” so skipped searches never claim to search. searchState distinguishes skipped, no_results, no_relevant_results, succeeded, and failed. Successful done events and non-streaming responses carry this state and optional sources; search failures carry searchState: "failed" in the error response. No usable evidence produces sources: [] and instructions to say that a search found insufficient relevant evidence, without fake citations. Initial retrieval failures use the existing error event (HTTP 503 on /chat) and do not generate an answer; an optional fallback failure preserves the successful initial search. The saved failed user message remains visible, and the composer restores its text for resending without search.

Sources appear below assistant messages as numbered links. They persist in a nullable SQLite chat_messages.sources JSON column, added automatically on startup using the existing additive schema mechanism; no manual migration is needed. Old messages remain valid. Sources are preserved in JSON exports/imports and branches. Regenerating an answer reuses its stored results without performing another search; a fresh search requires sending another message with the toggle enabled.

Run node --test --test-isolation=none and npm run lint to check the feature and existing flows. test/webSearch.test.js uses local HTTP protocol fixtures and an isolated in-memory database; no production search results are mocked. Live SearXNG/Ollama verification still requires configured, reachable services. There is no TypeScript or build command in this framework-free project.

Legal

• Privacy Policy
• Terms of Service

Page navigation

• / is the public landing page, even for signed-in visitors.
• /login and /register provide dedicated account forms. Registration starts a session automatically.
• /app is workspace home, with recent chats and optional starters.
• /app/chats/new opens an empty composer; sending the first message creates the conversation.
• /app/chats/:id opens a conversation. Legacy /?chat=:id links redirect here.
• /app/personas provides persona management, using the shared settings editor module.
• /app/settings contains account and appearance settings. /settings redirects here.
• /market allows visitors to browse public summaries and signed-in users to use the full persona market.

Protected pages redirect to sign-in with a local return destination. Public discovery only returns published persona summaries; using a persona requires an account.

All page routes render the same navigation from src/pageLayout.js into the <!-- site-header --> placeholder. public/shell.css owns the shared header, page spacing, and mobile navigation; public/app/workspaceNavigation.js owns sign-out. Page-specific headers contain only local titles and actions. The chat sidebar is reserved for conversation history. Public and home content share public/pages.css and public/pages.js. Page routes run before static assets and authenticated API routers; rendered pages use private, non-cacheable responses.

For environments that block child-process test workers, run node --test --test-isolation=none.