librechat.yaml

# =============================================================================
# Wiseway Staff Assistant — LibreChat runtime config
# =============================================================================
# Mounted into the container at /app/librechat.yaml (CONFIG_PATH points here).
# Keep ALL Wiseway customization here — never patch LibreChat source.
#
# Three branding knobs live in this repo:
#   1. APP_TITLE env (deploy/.env)        -> browser tab + app name
#   2. interface.customWelcome (below)    -> greeting on a fresh chat
#   3. interface.customFooter (below)     -> footer line under the composer
#   4. deploy/wiseway-assets/*.svg        -> logo + favicon (mounted over assets)
# See deploy/README.md for which knob does what.
# =============================================================================

version: 1.2.8
cache: true

# ---------------------------------------------------------------------------
# MCP settings — allowlist the internal doc-search hostname so LibreChat's
# SSRF guard permits the streamable-http connection. Without this, LibreChat
# blocks private/internal hostnames ("Domain ... is not allowed").
# ---------------------------------------------------------------------------
mcpSettings:
  allowedDomains:
    - "wiseway-doc-search"

# ---------------------------------------------------------------------------
# Interface — branding + which feature buttons are exposed
# ---------------------------------------------------------------------------
interface:
  # APP_TITLE is read from env; the title that shows in the tab / header.
  # (Set via deploy/.env -> APP_TITLE=Wiseway Assistant)
  customWelcome: "Welcome to the Wiseway Staff Assistant. Ask about HR policy, SOPs, or safety procedures — answers are drawn only from the documents your role is allowed to see, and every fact is cited."
  customFooter: "Wiseway Logistics — internal use only"
  # Agents endpoint is the host for the Wiseway HR & SOP Assistant agent.
  agents: true
  # No native file upload search or web search for this demo — the MCP
  # doc-search tool is the only retrieval path, and it is role-scoped.
  fileSearch: false
  webSearch: false

# ---------------------------------------------------------------------------
# Endpoints — local Ollama (native on the host, reached via host.docker.internal)
# ---------------------------------------------------------------------------
endpoints:
  # Agents endpoint — required so we can build an agent that uses the
  # role-scoped MCP doc-search tools. "tools" capability enables MCP tools.
  agents:
    disableBuilder: false
    capabilities:
      - "tools"
      - "actions"
      - "chain"
  custom:
    - name: "Ollama"
      # Ollama ignores the API key but LibreChat requires a non-empty value.
      apiKey: "ollama"
      # Ollama runs NATIVELY on the macOS host (Metal GPU), not in Docker.
      baseURL: "http://host.docker.internal:11434/v1/"
      models:
        default:
          - "qwen2.5:7b"
        # fetch:false — list models statically instead of querying Ollama at
        # request time. A live fetch is racy: if it hasn't resolved when an agent
        # initializes, model validation throws "endpoint_models_not_loaded" and
        # the user sees "Models for Ollama could not be loaded". Static = reliable.
        fetch: false
      titleConvo: true
      titleModel: "qwen2.5:7b"
      # Single-message title payloads work best as a "user" role for local models.
      titleMessageRole: "user"
      modelDisplayLabel: "Ollama (qwen2.5:7b)"

# ---------------------------------------------------------------------------
# MCP servers — role-scoped Wiseway document search
# ---------------------------------------------------------------------------
# Identity is injected ONLY via headers. The doc-search service trusts these
# because the request also carries the shared MCP key. Role-based filtering
# happens server-side in the MCP tool, BEFORE results reach the model.
#   X-User-Role -> {{LIBRECHAT_USER_ROLE}}  (warehouse | driver | office | hr-admin,
#                  seeded per-user in Mongo — see deploy/seed-roles.sh)
#   X-User-Id   -> {{LIBRECHAT_USER_ID}}
#   X-Mcp-Key   -> ${WISEWAY_MCP_SHARED_SECRET}  (env from deploy/.env)
mcpServers:
  wiseway-docs:
    type: streamable-http
    url: "http://wiseway-doc-search:8000/mcp"
    headers:
      X-User-Role: "{{LIBRECHAT_USER_ROLE}}"
      X-User-Id: "{{LIBRECHAT_USER_ID}}"
      X-Mcp-Key: "${WISEWAY_MCP_SHARED_SECRET}"
    timeout: 30000
    startup: true
    chatMenu: false

# ---------------------------------------------------------------------------
# Model specs — the curated "Wiseway HR & SOP Assistant" entry
# ---------------------------------------------------------------------------
# Two SOP agents are seeded automatically by deploy/create-agent.sh (via the
# LibreChat REST API), which also shares them publicly and patches the agent_ids
# below: "SOP Claude" (Anthropic Sonnet, the default for everyone) and "SOP"
# (local Ollama, the fallback). Both use the SAME gated wiseway-docs tools + the
# same citing prompt — only the model differs.
# The manual UI recipe is kept here for reference / customization (Agent Builder):
#   1. Sign in as an admin, open Agents -> create a new agent.
#   2. Endpoint = Ollama, model = qwen2.5:7b.
#   3. Add the MCP tools "search" and "fetch" from the wiseway-docs server.
#   4. Paste the system prompt below into the agent's instructions.
#   5. Save, copy the resulting agent id (looks like "agent_xxxxxxxx"), and
#      paste it into preset.agent_id below, then restart LibreChat.
#
# Example agent system prompt (the citing instruction is load-bearing — the
# local model will NOT auto-cite, so we instruct it explicitly):
#
#   You are the Wiseway HR & SOP Assistant. Answer staff questions about HR
#   policy, standard operating procedures, and workplace safety using ONLY the
#   wiseway-docs search and fetch tools. Always call search first. Cite every
#   fact as a Markdown link [Title](source_url) using the title and source_url
#   from each search hit. If search returns nothing relevant, say plainly that
#   you do not know and that the answer is not in the documents you can access.
#   Never invent policy. You only ever see documents this staff member's role
#   is permitted to read.
#
# Pin the Wiseway HR & SOP Assistant as the default landing experience and
# brand it with the Wiseway logo. iconURL points at the mounted Wiseway
# logo.svg (served from /assets/); agent_id is the real id from the Agent
# Builder. The agent's own avatar (set in Mongo) is what renders in-chat.
modelSpecs:
  prioritize: true
  # enforce:false lets staff fall back to the plain Ollama model — and pick any
  # OTHER agent the admin has shared — if the pinned SOP agent is unavailable.
  # *** Set enforce:true for production *** to LOCK staff to the gated SOP agent,
  # so they can never get un-cited, un-role-scoped answers from the bare model.
  # (A plain-model chat has NO wiseway-docs tool, so no role gate and no citations
  # — it can't leak docs, but it also can't honour the access policy.)
  enforce: false
  # Endpoints staff may pick in the CHAT selector. Deliberately NO cloud models
  # here: Anthropic/OpenAI/Google are admin-managed and must NOT be selectable
  # (or key-editable) by staff. They stay fully available to the admin in the
  # Agent Builder's Provider dropdown (which is independent of this list), so the
  # admin can still build agents on them. Staff are limited to the curated SOP
  # agent + the local Ollama model.
  #   NOTE: addedEndpoints is global (applies to every role). Cloud models are
  #   kept out of staff reach by (a) their absence here and (b) the ADMIN-only
  #   "create agents" permission set in the Admin Settings panel. interface
  #   `modelSelect`/`modelSpecs.enforce` are global toggles, so they can't hide
  #   models from staff-only — role-based agent permissions are the lever.
  addedEndpoints:
    - "agents"
    - "Ollama"
  list:
    # Default for everyone: the Claude (Anthropic Sonnet) SOP agent. Same gated
    # MCP tools + citing prompt as the local SOP agent — only the model differs.
    # Runs on endpoint "agents" (the agent calls Claude internally), so staff
    # never pick the raw Anthropic endpoint and the "no cloud models in the staff
    # selector" rule above still holds. Requires ANTHROPIC_API_KEY in deploy/.env.
    - name: "sop-claude"
      label: "SOP Claude"
      default: true
      description: "Cited answers on HR policy, SOPs and safety — scoped to your role."
      iconURL: "/assets/icon-192x192.png"
      greeting: "Hi — I'm the Wiseway HR & SOP Assistant. Ask me about leave, rosters, safety procedures, or any SOP. I only answer from the documents your role can see, and I link the source for every answer."
      preset:
        endpoint: "agents"
        # Seeded + auto-patched by deploy/create-agent.sh (the "SOP Claude" agent).
        agent_id: "agent_pE30gepVMJyNV3kHm6-Q4"
    # The local (Ollama / qwen2.5:7b) SOP agent — kept as a no-cloud fallback that
    # staff can still pick (enforce:false). Identical role gate + citing prompt.
    - name: "sop"
      label: "SOP (local)"
      description: "Cited answers on HR policy, SOPs and safety — scoped to your role (local model)."
      iconURL: "/assets/icon-192x192.png"
      greeting: "Hi — I'm the Wiseway HR & SOP Assistant. Ask me about leave, rosters, safety procedures, or any SOP. I only answer from the documents your role can see, and I link the source for every answer."
      preset:
        endpoint: "agents"
        # Seeded + auto-patched by deploy/create-agent.sh (the "SOP" agent).
        agent_id: "agent_KfDm7NcIG3gia3vf8UfQS"