docs: add docker.md — full Docker distribution architecture plan

Covers all six services (jarvisChat, SearXNG, Qdrant, RabbitMQ,
llama-server, Ollama) with images, ports, volumes, healthchecks,
env mapping, secrets management, setup wizard spec, and back-out
procedure. Includes pre-v1.0 gate checklist and open decisions.

Also add .env, secrets/, searxng/ to .gitignore (these are
generated by setup.sh and contain secrets).

.gitignore

@@ -6,3 +6,6 @@ venv/
 readme.md-
 *.bak
 hardware_state.json
+.env
+secrets/
+searxng/

app.py

@@ -14,7 +14,7 @@ from fastapi.responses import HTMLResponse, JSONResponse
 from fastapi.staticfiles import StaticFiles
 from fastapi.templating import Jinja2Templates
 
-from config import VERSION, RATE_WINDOW_SECONDS, UPLOAD_DIR
+from config import VERSION, RATE_WINDOW_SECONDS, UPLOAD_DIR, RAG_MAX_VECTORS, RAG_EVICTION_HIGH_WATER, RAG_EVICTION_LOW_WATER, RAG_EVICTION_BATCH
 from db import init_db
 from hardware import assess_hardware
 from memory import get_memory_count
@@ -37,6 +37,7 @@ import routers.completions as completions
 import routers.upload as upload
 import routers.ingest as ingest
 import routers.hardware as hardware
+import routers.rag_admin as rag_admin
 
 # --- Logging ---
 log = logging.getLogger("jarvischat")
@@ -56,6 +57,18 @@ async def lifespan(app: FastAPI):
     init_db()
     log.info(f"Memory system: {get_memory_count()} memories loaded")
     await assess_hardware()
+
+    if RAG_MAX_VECTORS > 0:
+        if RAG_EVICTION_HIGH_WATER <= RAG_EVICTION_LOW_WATER:
+            log.warning(
+                f"RAG_EVICTION_HIGH_WATER={RAG_EVICTION_HIGH_WATER} <= "
+                f"RAG_EVICTION_LOW_WATER={RAG_EVICTION_LOW_WATER} — eviction will never fire"
+            )
+        if RAG_EVICTION_BATCH <= 0:
+            log.warning(f"RAG_EVICTION_BATCH={RAG_EVICTION_BATCH} clamped to 1")
+    else:
+        log.warning("RAG_MAX_VECTORS <= 0 — RAG eviction disabled")
+
     yield
     log.info("JarvisChat shutting down")
 
@@ -146,6 +159,7 @@ for router_module in [
     auth_router, conversations.router, memories.router, models.router,
     presets.router, profile.router, settings.router, skills.router,
     chat.router, search_route.router, completions.router, upload.router, ingest.router, hardware.router,
+    rag_admin.router,
 ]:
     app.include_router(router_module)
 

config.py

@@ -9,7 +9,7 @@ import logging
 
 log = logging.getLogger("jarvischat")
 
-VERSION = "v0.11.0"
+VERSION = "v0.13.0"
 OLLAMA_BASE = os.environ.get("OLLAMA_BASE", "http://localhost:11434")
 LLAMA_SERVER_BASE = os.environ.get("LLAMA_SERVER_BASE", "http://192.168.50.108:8081")
 SEARXNG_BASE = "http://localhost:8888"
@@ -50,9 +50,21 @@ BODY_LIMIT_PROFILE_BYTES = 256 * 1024
 UPLOAD_DIR = "/tmp/jarvischat_uploads"
 MAX_UPLOAD_BYTES = 20 * 1024 * 1024
 SUPPORTED_UPLOAD_TYPES = {"text/plain", "text/markdown", "application/pdf", "application/json", "text/x-python", "text/html", "image/png", "image/jpeg", "image/gif", "image/svg+xml", "image/webp"}
+QDRANT_URL = "http://192.168.50.108:6333"
+RAG_COLLECTION = "jarvis_rag"
 UPLOAD_CONTEXT_EXPIRY_HOURS = 1
 BODY_LIMIT_UPLOAD_BYTES = MAX_UPLOAD_BYTES
 
+# --- RAG eviction ---
+RAG_MAX_VECTORS = 50000
+RAG_EVICTION_HIGH_WATER = 0.80
+RAG_EVICTION_LOW_WATER = 0.20
+RAG_EVICTION_BATCH = 1000
+RAG_PINNED_SOURCES = ["upload", "profile"]
+RAG_GRACE_HOURS = 1
+RAG_ACCESS_WEIGHT = 1.0
+RAG_AGE_WEIGHT = 0.1
+
 MAX_CHAT_MESSAGE_CHARS = 8000
 MAX_SEARCH_QUERY_CHARS = 500
 MAX_PROFILE_CHARS = 32000

docker.md

@@ -0,0 +1,681 @@
+# Docker Distribution — Architecture & Planning
+
+> **Part of B3 (v1.0 gate).** This document catalogs every service, volume, port, configuration, and decision needed to ship jarvisChat as a `docker compose` stack. It also defines extraction (setup) and back-out (uninstall) procedures so nothing is lost when reality disagrees with the plan.
+
+## 1. Stack Overview
+
+```
+┌─────────────────────────────────────────────────────────┐
+│                    docker compose stack                  │
+│                                                         │
+│  ┌────────────┐  ┌──────────┐  ┌────────────────────┐  │
+│  │  SearXNG    │  │  Qdrant  │  │  RabbitMQ          │  │
+│  │  :8888      │  │  :6333   │  │  :5672 / :15672    │  │
+│  └──────┬──────┘  └────┬─────┘  └────────┬───────────┘  │
+│         │              │                  │              │
+│         ▼              ▼                  ▼              │
+│  ┌──────────────────────────────────────────────────┐   │
+│  │              jarvisChat (FastAPI)                │   │
+│  │              :8080 (HTTP)                        │   │
+│  │                                                  │   │
+│  │  SQLite  ◄── jarvischat.db           (volume)    │   │
+│  │  Uploads ◄── /app/uploads            (volume)    │   │
+│  └──────────┬──────────────┬───────────────────────┘   │
+│             │              │                            │
+│             ▼              ▼                            │
+│  ┌──────────────┐  ┌──────────────┐                    │
+│  │ llama-server │  │   Ollama     │                    │
+│  │ :8081        │  │ :11434       │                    │
+│  │ (GPU/RPC)    │  │ (embeddings) │                    │
+│  └──────────────┘  └──────────────┘                    │
+└─────────────────────────────────────────────────────────┘
+```
+
+### Service roles
+
+| Service | Image | Role |
+|---------|-------|------|
+| **jarvisChat** | Custom `Dockerfile` | FastAPI app serving UI + API |
+| **SearXNG** | `searxng/searxng:latest` | Privacy-respecting web search |
+| **Qdrant** | `qdrant/qdrant:latest` | Vector database for RAG |
+| **RabbitMQ** | `rabbitmq:4-management` | Message broker for AMQP cluster |
+| **llama-server** | `ghcr.io/ggml-org/llama.cpp:server` | LLM inference (OpenAI-compat API) |
+| **Ollama** | `ollama/ollama:latest` | Embeddings for RAG chunk vectors |
+
+### Non-containerized (host-level)
+
+| Component | Reason |
+|-----------|--------|
+| AMD GPU driver + ROCm | Kernel access required for GPU compute |
+| llama.cpp RPC workers | Runs on *other* hosts — not on the Docker host |
+| `rocm-smi` | Hardware stats — not needed for core function |
+| `psutil` | Already inside the container via pip |
+
+---
+
+## 2. Service Catalog
+
+### 2.1 jarvisChat (FastAPI app)
+
+**Image:** `jarvischat:latest` (built from `Dockerfile`)
+
+**Ports:**
+| Container | Host | Purpose |
+|-----------|------|---------|
+| 8080 | 8080 | HTTP API + UI |
+
+**Volumes:**
+| Container path | Type | Purpose |
+|----------------|------|---------|
+| `/app/jarvischat.db` | named volume `jarvischat_data` | SQLite database |
+| `/app/uploads` | named volume `jarvischat_uploads` | Uploaded files |
+| `/app/hardware_state.json` | (inside volume) | Cached hardware probe |
+
+**Dependencies:** Wait for SearXNG, Qdrant, RabbitMQ, llama-server, Ollama before serving.
+
+**Restart:** `unless-stopped`
+
+**Healthcheck:** `curl -f http://localhost:8080/`
+
+### 2.2 SearXNG
+
+**Image:** `searxng/searxng:latest`
+
+**Ports:**
+| Container | Host | Purpose |
+|-----------|------|---------|
+| 8080 | 8888 | Search API |
+
+**Volumes:**
+| Container path | Type | Purpose |
+|----------------|------|---------|
+| `/etc/searxng` | named volume `searxng_config` | `settings.yml` |
+
+**Environment:**
+```env
+SEARXNG_BASE_URL=https://localhost:8888
+```
+
+**Config override (`/etc/searxng/settings.yml`):**
+```yaml
+search:
+  safe_search: 0
+  autocomplete: ""
+server:
+  secret_key: ${SEARXNG_SECRET_KEY}
+  limiter: false
+  image_proxy: false
+  method: GET
+  port: 8080
+  bind_address: "0.0.0.0"
+```
+
+**Restart:** `unless-stopped`
+
+### 2.3 Qdrant
+
+**Image:** `qdrant/qdrant:latest`
+
+**Ports:**
+| Container | Host | Purpose |
+|-----------|------|---------|
+| 6333 | 6333 | HTTP API |
+| 6334 | — | gRPC (internal only) |
+
+**Volumes:**
+| Container path | Type | Purpose |
+|----------------|------|---------|
+| `/qdrant/storage` | named volume `qdrant_storage` | Vector index data |
+
+**Environment:**
+```env
+QDRANT__SERVICE__GRPC_PORT=6334
+```
+
+**Restart:** `unless-stopped`
+
+### 2.4 RabbitMQ
+
+**Image:** `rabbitmq:4-management`
+
+**Ports:**
+| Container | Host | Purpose |
+|-----------|------|---------|
+| 5672 | 5672 | AMQP messaging |
+| 15672 | — | Management UI (internal only) |
+
+**Volumes:**
+| Container path | Type | Purpose |
+|----------------|------|---------|
+| `/var/lib/rabbitmq` | named volume `rabbitmq_data` | Message store |
+
+**Environment:**
+```env
+RABBITMQ_DEFAULT_USER=jarvischat
+RABBITMQ_DEFAULT_PASS_FILE=/run/secrets/rabbitmq_password
+RABBITMQ_DEFAULT_VHOST=/
+```
+
+**Restart:** `unless-stopped`
+
+### 2.5 llama-server
+
+**Image:** `ghcr.io/ggml-org/llama.cpp:server`
+
+**Ports:**
+| Container | Host | Purpose |
+|-----------|------|---------|
+| 8081 | 8081 | OpenAI-compat API |
+
+**Volumes:**
+| Container path | Type | Purpose |
+|----------------|------|---------|
+| `/models` | bind mount `./models` | Model GGUF files |
+
+**Environment:**
+```env
+LLAMA_ARG_MODEL=/models/<model-file>
+LLAMA_ARG_N_GPU_LAYERS=0              # set >0 for GPU offload
+LLAMA_ARG_MAIN_GPU=0
+LLAMA_ARG_CTX_SIZE=4096
+LLAMA_ARG_HOST=0.0.0.0
+LLAMA_ARG_PORT=8081
+LLAMA_ARG_EMBEDDINGS=1
+LLAMA_ARG_LOGPROBS=1
+LLAMA_ARG_RPC=                         # optional: comma-separated RPC endpoints
+```
+
+**Restart:** `unless-stopped`
+
+**Healthcheck:** `curl -f http://localhost:8081/health`
+
+**Notes:**
+- Models directory bind mount — user places `.gguf` files in `./models/` on the host
+- RPC offload to other machines (e.g., `10.0.0.50:50052,10.0.0.51:50052`)
+- If no GPU, set `LLAMA_ARG_N_GPU_LAYERS=0` for CPU-only
+- `LLAMA_ARG_EMBEDDINGS=1` required for perplexity scoring
+- `LLAMA_ARG_LOGPROBS=1` required for auto-search trigger
+
+### 2.6 Ollama
+
+**Image:** `ollama/ollama:latest`
+
+**Ports:**
+| Container | Host | Purpose |
+|-----------|------|---------|
+| 11434 | 11434 | Embeddings API |
+
+**Volumes:**
+| Container path | Type | Purpose |
+|----------------|------|---------|
+| `/root/.ollama` | named volume `ollama_models` | Pulled model blobs |
+
+**Restart:** `unless-stopped`
+
+**Notes:**
+- Used exclusively for embeddings (`/api/embeddings`), not inference
+- Typically needs a small model like `all-minilm:latest` or `nomic-embed-text:latest`
+- Consider replacing Ollama with llama-server's built-in embedding if it supports the same model — would remove one container
+
+---
+
+## 3. Configuration Management
+
+### 3.1 `.env` file (generated by setup wizard)
+
+```env
+# --- Secrets (auto-generated, change before production) ---
+JARVISCHAT_ADMIN_PIN=
+JARVISCHAT_COMPLETIONS_API_KEY=
+JARVISCHAT_ALLOW_DEFAULT_PIN=false
+RABBITMQ_PASSWORD=
+SEARXNG_SECRET_KEY=
+
+# --- Host discovery (auto-detected by setup wizard) ---
+LLAMA_SERVER_BASE=http://llama-server:8081
+OLLAMA_BASE=http://ollama:11434
+SEARXNG_BASE=http://searxng:8888
+QDRANT_URL=http://qdrant:6333
+RABBITMQ_HOST=rabbitmq
+RABBITMQ_PORT=5672
+
+# --- Performance tuning (calculated by setup wizard) ---
+RAG_MAX_VECTORS=50000
+RAG_EVICTION_HIGH_WATER=0.80
+RAG_EVICTION_LOW_WATER=0.20
+RAG_EVICTION_BATCH=1000
+
+# --- llama-server options ---
+LLAMA_MODEL=llama3.1-8b-instruct.Q4_K_M.gguf
+LLAMA_N_GPU_LAYERS=0
+LLAMA_RPC_ENDPOINTS=
+LLAMA_CTX_SIZE=4096
+
+# --- Ollama ---
+OLLAMA_EMBED_MODEL=all-minilm:latest
+
+# --- Network ---
+JARVISCHAT_ALLOWED_CIDRS=127.0.0.0/8,::1/128,10.0.0.0/8,172.16.0.0/12,192.168.0.0/16
+JARVISCHAT_TRUSTED_ORIGINS=
+JARVISCHAT_TRUST_X_FORWARDED_FOR=false
+```
+
+### 3.2 Mapping of config.py → .env variable
+
+Every config.py default that references an external service must accept a matching env var at runtime:
+
+| config.py constant | .env variable | Service |
+|-------------------|---------------|---------|
+| `LLAMA_SERVER_BASE` | `LLAMA_SERVER_BASE` | llama-server |
+| `OLLAMA_BASE` | `OLLAMA_BASE` | Ollama |
+| `SEARXNG_BASE` | `SEARXNG_BASE` | SearXNG |
+| `QDRANT_URL` | `QDRANT_URL` | Qdrant |
+| `COMPLETIONS_API_KEY` | `JARVISCHAT_COMPLETIONS_API_KEY` | — |
+| `ALLOWED_CIDRS_RAW` | `JARVISCHAT_ALLOWED_CIDRS` | — |
+| `TRUST_X_FORWARDED_FOR` | `JARVISCHAT_TRUST_X_FORWARDED_FOR` | — |
+| `TRUSTED_ORIGINS` | `JARVISCHAT_TRUSTED_ORIGINS` | — |
+| `RAG_MAX_VECTORS` | `RAG_MAX_VECTORS` | — (calc'd from RAM) |
+
+### 3.3 Secrets management
+
+| Secret | Generated by | Stored in | Mounted to |
+|--------|-------------|-----------|------------|
+| `JARVISCHAT_ADMIN_PIN` | User prompt | `.env` | jarvisChat container |
+| `JARVISCHAT_COMPLETIONS_API_KEY` | Auto-generated, shown to user | `.env` | jarvisChat container |
+| `RABBITMQ_PASSWORD` | Auto-generated | `.env` + Docker secret | RabbitMQ container |
+| `SEARXNG_SECRET_KEY` | Auto-generated | `.env` | SearXNG container |
+
+**Docker secrets approach:** Use `secrets:` in compose file for RabbitMQ password (mounted as file) rather than passing via env var, since `settings.yml` in SearXNG and RabbitMQ config can reference file-based secrets without env-var leakage.
+
+### 3.4 Dockerfile for jarvisChat
+
+```dockerfile
+FROM python:3.13-slim-bookworm AS builder
+WORKDIR /app
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+FROM python:3.13-slim-bookworm
+WORKDIR /app
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    curl && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=builder /usr/local/lib/python3.13/site-packages /usr/local/lib/python3.13/site-packages
+COPY --from=builder /usr/local/bin /usr/local/bin
+COPY . .
+
+EXPOSE 8080
+HEALTHCHECK --interval=30s --timeout=10s --start-period=15s --retries=3 \
+    CMD curl -f http://localhost:8080/ || exit 1
+
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "8080"]
+```
+
+**Multi-stage rationale:** First stage compiles/bundles packages (wheels), final stage is minimal. Devs can skip builder with `--target builder` for live-reload with volume mount.
+
+---
+
+## 4. docker-compose.yml structure
+
+```yaml
+services:
+  jarvischat:
+    build: .
+    ports: ["8080:8080"]
+    volumes:
+      - jarvischat_data:/app/jarvischat.db
+      - jarvischat_uploads:/app/uploads
+    env_file: .env
+    depends_on:
+      searxng: { condition: service_started }
+      qdrant: { condition: service_started }
+      rabbitmq: { condition: service_healthy }
+      llama-server: { condition: service_healthy }
+      ollama: { condition: service_started }
+    restart: unless-stopped
+
+  searxng:
+    image: searxng/searxng:latest
+    ports: ["8888:8080"]
+    volumes:
+      - ./searxng/settings.yml:/etc/searxng/settings.yml:ro
+      - searxng_config:/etc/searxng
+    env_file: .env
+    restart: unless-stopped
+
+  qdrant:
+    image: qdrant/qdrant:latest
+    ports: ["6333:6333"]
+    volumes:
+      - qdrant_storage:/qdrant/storage
+    restart: unless-stopped
+
+  rabbitmq:
+    image: rabbitmq:4-management
+    ports: ["5672:5672"]
+    volumes:
+      - rabbitmq_data:/var/lib/rabbitmq
+    env_file: .env
+    secrets:
+      - rabbitmq_password
+    healthcheck:
+      test: ["CMD", "rabbitmq-diagnostics", "check_port_connectivity"]
+      interval: 15s
+      timeout: 5s
+      retries: 3
+    restart: unless-stopped
+
+  llama-server:
+    image: ghcr.io/ggml-org/llama.cpp:server
+    ports: ["8081:8081"]
+    volumes:
+      - ./models:/models:ro
+    env_file: .env
+    command: >
+      --model /models/${LLAMA_MODEL}
+      --host 0.0.0.0 --port 8081
+      --ctx-size ${LLAMA_CTX_SIZE:-4096}
+      --n-gpu-layers ${LLAMA_N_GPU_LAYERS:-0}
+      --embeddings
+      --logprobs
+      ${LLAMA_RPC_ENDPOINTS:+--rpc ${LLAMA_RPC_ENDPOINTS}}
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8081/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 5
+      start_period: 60s
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    restart: unless-stopped
+
+  ollama:
+    image: ollama/ollama:latest
+    ports: ["11434:11434"]
+    volumes:
+      - ollama_models:/root/.ollama
+    healthcheck:
+      test: ["CMD", "ollama", "list"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    restart: unless-stopped
+
+volumes:
+  jarvischat_data:
+  jarvischat_uploads:
+  searxng_config:
+  qdrant_storage:
+  rabbitmq_data:
+  ollama_models:
+
+secrets:
+  rabbitmq_password:
+    file: ./secrets/rabbitmq_password.txt
+```
+
+**Notes:**
+- GPU reservations use `resources.reservations.devices` — this is compose v3.8+. For AMD GPUs, replace `driver: nvidia` with `driver: amd` (experimental Docker support). For hosts without GPU, omit the `deploy` block entirely.
+- The `deploy` block only applies when deployed as a swarm stack. For `docker compose`, GPU access may need `--gpus all` or `device_requests` in config. Verify compatibility.
+- SearXNG config file (`settings.yml`) is bind-mounted read-only from the host repo clone — the setup wizard should generate this file.
+
+---
+
+## 5. Networking
+
+### 5.1 Internal communication (compose network)
+
+| From | To | Port | Protocol |
+|------|----|------|----------|
+| jarvisChat | llama-server | 8081 | HTTP |
+| jarvisChat | Ollama | 11434 | HTTP |
+| jarvisChat | SearXNG | 8080 | HTTP |
+| jarvisChat | Qdrant | 6333 | HTTP |
+| jarvisChat | RabbitMQ | 5672 | AMQP |
+| RabbitMQ | (cluster peers) | 4369 | EPMD |
+| RabbitMQ | (cluster peers) | 25672 | Inter-node |
+
+### 5.2 Exposed ports (host-facing)
+
+| Port | Service | Should expose? | Notes |
+|------|---------|---------------|-------|
+| 8080 | jarvisChat | ✅ Required | UI + API |
+| 8888 | SearXNG | Optional | Only if user wants standalone search |
+| 6333 | Qdrant | Optional | Only for external tooling |
+| 5672 | RabbitMQ | Optional | Only for remote AMQP clients |
+| 15672 | RabbitMQ mgmt | ❌ Internal | Healthcheck only |
+| 8081 | llama-server | Optional | Only for external tooling |
+| 11434 | Ollama | Optional | Only for external tooling |
+
+**Design decision:** By default, only port 8080 (jarvisChat) is published. All other services remain on the internal compose network. Advanced users can opt-in by uncommenting `ports:` blocks.
+
+### 5.3 Reverse proxy consideration
+
+For production, a reverse proxy (Caddy, nginx, Traefik) should sit in front:
+
+```yaml
+# Optional — compose profile: "proxy"
+caddy:
+  image: caddy:latest
+  ports: ["80:80", "443:443"]
+  volumes:
+    - ./Caddyfile:/etc/caddy/Caddyfile:ro
+    - caddy_data:/data
+```
+
+This is out of scope for v1.0 but documented for future.
+
+---
+
+## 6. Setup Wizard (Extraction)
+
+`setup.sh` — idempotent, interactive, runs on first boot.
+
+### Flow
+
+```
+1. CHECK: Is .env present?
+   ├── YES → skip to step 7 (or ask to regenerate)
+   └── NO  → continue
+
+2. INTRO: Print banner, explain what's about to happen
+
+3. PROBE: Run hardware assessment
+   ├── psutil → RAM total, CPU count
+   ├── rocm-smi → VRAM (optional, best-effort)
+   └── nvidia-smi → VRAM (optional, best-effort)
+
+4. NETWORK: Ask for
+   ├── Hostname / LAN IP for this machine
+   ├── Admin PIN (4 digits, or accept auto-generated)
+   └── (Optional) RPC endpoints for GPU offload
+
+5. CALCULATE:
+   ├── RAG_MAX_VECTORS = max(1000, int(available_ram_gb * 100_000))
+   ├── LLAMA_N_GPU_LAYERS = 0 (CPU default; offer GPU detection)
+   ├── LLAMA_MODEL = default gguf filename
+   └── RABBITMQ_PASSWORD = openssl rand -hex 20
+
+6. GENERATE:
+   ├── .env file from template
+   ├── ./secrets/rabbitmq_password.txt
+   ├── ./searxng/settings.yml (with generated secret_key)
+   └── ./models/README.txt (instructions for placing .gguf)
+
+7. VERIFY:
+   ├── docker and docker compose plugin installed
+   ├── docker compose version >= 2.x
+   ├── SUCCESS → "Run: docker compose up -d"
+   └── FAILURE  → show diagnostics and links
+
+8. EXTRACT model:
+   ├── Prompt for download URL or local path
+   ├── Offer to pull from HuggingFace if huggingface-cli available
+   └── Guides user to place file in ./models/
+```
+
+### What setup.sh creates on disk
+
+```
+./docker-deploy/
+├── .env                        # All env vars (SECRET — add to .gitignore)
+├── docker-compose.yml          # Compose stack definition
+├── Dockerfile                  # jarvisChat image build
+├── secrets/
+│   └── rabbitmq_password.txt   # RabbitMQ password file
+├── searxng/
+│   └── settings.yml            # SearXNG config with generated secret_key
+├── models/
+│   ├── README.txt              # Instructions for model placement
+│   └── <model>.gguf            # (user-provided)
+└── setup.log                   # Wizard run log
+```
+
+### Idempotency
+
+Re-running `setup.sh`:
+- With `.env` present: ask "Regenerate? This will overwrite existing config."
+- Without `.env`: fresh run
+- Never overwrites `./models/*.gguf` files
+- Never touches running containers — only modifies files on disk
+
+---
+
+## 7. Back-out Procedure (Uninstall)
+
+`teardown.sh` — returns the host system to its pre-install state.
+
+### What gets removed
+
+| Item | Removal method |
+|------|---------------|
+| Docker containers | `docker compose down -v` |
+| Docker images | `docker rmi jarvischat:latest` (ask about other images) |
+| Docker volumes | `docker volume rm jarvischat_data ...` (prompt first) |
+| Network `jarvischat_default` | Removed with compose |
+| `.env` file | `rm .env` |
+| `secrets/` directory | `rm -rf secrets/` |
+| `searxng/` directory | `rm -rf searxng/` |
+| `setup.log` | `rm setup.log` |
+| `hardware_state.json` | `rm hardware_state.json` |
+
+### What is preserved (by default)
+
+| Item | Reason |
+|------|--------|
+| `./models/*.gguf` | User data — prompt for deletion |
+| `jarvischat.db` (in volume) | Prompt: "Keep database snapshot?" |
+| `./uploads/` (in volume) | Prompt: "Keep uploaded files?" |
+| Docker Engine itself | Not installed by this project — leave it |
+
+### Script flow
+
+```
+1. CHECK: docker compose file exists?
+   ├── NO  → warn, continue
+   └── YES → docker compose down -v
+
+2. CHECK: .env exists?
+   ├── NO  → skip
+   └── YES → ask: "Remove .env?" (default no)
+
+3. ASK: "Remove secrets/ and searxng/ directories?" (default no)
+
+4. ASK: "Remove Docker images? (y/N)" (default no)
+   ├── Y → docker rmi jarvischat:latest
+   ├── Y → docker image ls | grep searxng/qdrant/rabbitmq → prompt per image
+   └── N → skip
+
+5. ASK: "Keep database volume snapshot? (Y/n)" (default yes)
+   ├── N → docker volume rm jarvischat_data
+   └── Y → leave volume (can be reattached later)
+
+6. ASK: "Remove model files from ./models/? (y/N)" (default no)
+
+7. CLEANUP generated artifacts:
+   ├── rm -f setup.log
+   ├── rm -f hardware_state.json
+   └── rm -f docker-compose.yml
+
+8. SUMMARY:
+   ├── "Docker stack removed"
+   ├── "Persistent data preserved at: <paths>"
+   └── "Models kept at: ./models/"
+```
+
+### Partial rollback
+
+If the setup wizard fails mid-way, a partial rollback is better than leaving detritus:
+
+| Failure point | Clean up |
+|--------------|----------|
+| After .env, before compose | `rm .env; rm -rf secrets/ searxng/` |
+| After compose, before first `up` | `rm docker-compose.yml; rm -rf *` |
+| After `up` but before healthcheck | `docker compose down -v; rm -rf ./*` |
+
+`setup.sh` should trap EXIT on failure and prompt: "Clean up partial install? [y/N]"
+
+---
+
+## 8. Open Decisions
+
+| Decision | Options | Priority |
+|----------|---------|----------|
+| **Ollama vs llama-server embeddings** | Both work. Keep both for now — remove Ollama if llama-server handles embeddings. Reduce containers = simpler. | Medium |
+| **GPU support in compose** | NVIDIA: well-supported. AMD: requires `--device=/dev/kfd --device=/dev/dri` and ROCm image. Document both. | High |
+| **RabbitMQ clustering vs single node** | Single node in v1.0. Clustering docs for multi-host later. | Low |
+| **SearXNG config management** | Bind-mount a generated `settings.yml`, or let container create default and post-process. Bind-mount is cleaner. | Medium |
+| **Reverse proxy** | Caddy is simplest for auto-HTTPS. Out of scope for v1.0 but design for it. | Low |
+| **Healthcheck strategy** | `depends_on` with `condition: service_healthy` is the safest approach but increases startup time. Acceptable. | Medium |
+| **Database migration** | SQLite file in volume — no migration needed for v1.0 format. If schema changes post-v1.0, need a migration container. | Low |
+| **Linux vs macOS vs Windows** | Linux-primary. macOS may work with changes (no rocm-smi). Windows via WSL2 only. | Low |
+| **LLM model download** | HuggingFace CLI integration in setup.sh, or manual download. Manual is simpler. | Low |
+| **Dockerfile optimization** | Pin pip hashes, use `--no-cache-dir`, consider `slim` vs `alpine`. Alpine has musl compatibility issues with psutil. Stay with slim. | Medium |
+
+## 9. Checklist (pre-v1.0 gate)
+
+- [ ] `Dockerfile` written and builds clean
+- [ ] `docker-compose.yml` boots all containers
+- [ ] jarvisChat container reaches all services (env vars resolve correctly)
+- [ ] SearXNG settings.yml generated correctly by setup.sh
+- [ ] RabbitMQ password secret mounted correctly
+- [ ] GPU (NVIDIA) passes through to llama-server container
+- [ ] GPU (AMD) passes through to llama-server container (or documented limitation)
+- [ ] `.env.example` checked in (no real secrets)
+- [ ] `setup.sh` written, idempotent, tested on clean Debian
+- [ ] `teardown.sh` written, tested, doesn't delete models without confirmation
+- [ ] `docker compose up -d` works without any manual steps beyond setup.sh
+- [ ] `docker compose down -v` followed by `setup.sh && docker compose up -d` = fresh stack
+- [ ] Healthchecks prevent serving before dependencies are ready
+- [ ] v1.0 release tag created
+
+---
+
+## 10. Files to create for B3
+
+```
+docker.md             ← this file (planning doc)
+Dockerfile            ← jarvisChat image
+docker-compose.yml    ← full stack
+.env.example          ← template without secrets
+setup.sh              ← extraction wizard
+teardown.sh           ← back-out utility
+searxng/
+  settings.yml        ← SearXNG config (generated by setup.sh)
+secrets/
+  rabbitmq_password.txt  ← generated by setup.sh
+models/
+  README.txt          ← instructions for placing .gguf
+```

eviction.py

@@ -0,0 +1,268 @@
+"""
+JarvisChat — Score-based RAG vector eviction with hysteresis.
+"""
+import asyncio
+import logging
+from datetime import datetime, timezone, timedelta
+
+import httpx
+
+from config import (
+    QDRANT_URL, RAG_COLLECTION,
+    RAG_MAX_VECTORS, RAG_EVICTION_HIGH_WATER, RAG_EVICTION_LOW_WATER,
+    RAG_EVICTION_BATCH, RAG_PINNED_SOURCES, RAG_GRACE_HOURS,
+    RAG_ACCESS_WEIGHT, RAG_AGE_WEIGHT,
+)
+
+log = logging.getLogger("jarvischat")
+
+eviction_lock = asyncio.Lock()
+EVICTION_LOG: list[dict] = []
+
+
+async def _update_retrieval_count(point_id: str, current_count: int = 0):
+    try:
+        async with httpx.AsyncClient() as client:
+            payload = {
+                "retrieval_count": current_count + 1,
+                "last_accessed": datetime.now(timezone.utc).isoformat(),
+            }
+            resp = await client.put(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}/points/payload",
+                json={"points": [point_id], "payload": payload},
+                timeout=5.0,
+            )
+            if resp.status_code not in (200, 201):
+                log.warning(f"Failed to increment retrieval count for {point_id}: {resp.status_code}")
+    except Exception as e:
+        log.warning(f"Error incrementing retrieval count for {point_id}: {e}")
+
+
+async def get_collection_count() -> int:
+    try:
+        async with httpx.AsyncClient() as client:
+            resp = await client.get(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}",
+                timeout=10.0,
+            )
+            if resp.status_code == 200:
+                return resp.json().get("result", {}).get("vectors_count", 0)
+    except Exception as e:
+        log.warning(f"get_collection_count error: {e}")
+    return 0
+
+
+async def get_collection_stats() -> dict:
+    count = await get_collection_count()
+    high_water_pct = int(RAG_EVICTION_HIGH_WATER * 100)
+    low_water_pct = int(RAG_EVICTION_LOW_WATER * 100)
+    percent_full = round((count / RAG_MAX_VECTORS) * 100, 1) if RAG_MAX_VECTORS > 0 else 0
+    return {
+        "vector_count": count,
+        "max_vectors": RAG_MAX_VECTORS,
+        "high_water_mark": int(RAG_MAX_VECTORS * RAG_EVICTION_HIGH_WATER),
+        "low_water_mark": int(RAG_MAX_VECTORS * RAG_EVICTION_LOW_WATER),
+        "high_water_pct": high_water_pct,
+        "low_water_pct": low_water_pct,
+        "percent_full": percent_full,
+        "pinned_sources": list(RAG_PINNED_SOURCES),
+    }
+
+
+async def evict_batch(batch_size: int) -> int:
+    filter_conditions = {
+        "must_not": [
+            {"match": {"key": "source", "value": src}}
+            for src in RAG_PINNED_SOURCES
+        ]
+    }
+    try:
+        async with httpx.AsyncClient() as client:
+            scroll_resp = await client.post(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}/points/scroll",
+                json={
+                    "filter": filter_conditions,
+                    "limit": min(batch_size * 10, 10000),
+                    "with_payload": True,
+                    "with_vector": False,
+                },
+                timeout=30.0,
+            )
+            if scroll_resp.status_code != 200:
+                log.warning(f"Eviction scroll failed: {scroll_resp.status_code}")
+                return 0
+
+            points = scroll_resp.json().get("result", {}).get("points", [])
+            if not points:
+                return 0
+
+            now = datetime.now(timezone.utc)
+            scored = []
+            for p in points:
+                payload = p.get("payload", {})
+                date_str = payload.get("ingest_date") or payload.get("upload_date", "")
+                if date_str:
+                    age_hours = (now - datetime.fromisoformat(date_str)).total_seconds() / 3600
+                else:
+                    age_hours = 999999
+
+                if age_hours < RAG_GRACE_HOURS:
+                    continue
+
+                retrieval_count = payload.get("retrieval_count", 0) or 0
+                score = retrieval_count * RAG_ACCESS_WEIGHT + age_hours * RAG_AGE_WEIGHT
+                last_accessed = payload.get("last_accessed", date_str)
+                scored.append((score, last_accessed, p["id"]))
+
+            if not scored:
+                log.warning("No evictable vectors found (all pinned or newborn)")
+                return 0
+
+            scored.sort(key=lambda x: (x[0], x[1]))
+            to_delete = [p[2] for p in scored[:batch_size]]
+            if not to_delete:
+                return 0
+
+            delete_resp = await client.post(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}/points/delete",
+                json={"points": to_delete},
+                timeout=30.0,
+            )
+            if delete_resp.status_code not in (200, 201):
+                log.warning(f"Eviction delete failed: {delete_resp.status_code}")
+                return 0
+
+            return len(to_delete)
+    except Exception as e:
+        log.warning(f"evict_batch error: {e}")
+        return 0
+
+
+async def maybe_evict() -> int:
+    if RAG_MAX_VECTORS <= 0:
+        return 0
+    effective_batch = max(RAG_EVICTION_BATCH, 1)
+
+    async with eviction_lock:
+        count = await get_collection_count()
+        threshold_high = int(RAG_MAX_VECTORS * RAG_EVICTION_HIGH_WATER)
+        threshold_low = int(RAG_MAX_VECTORS * RAG_EVICTION_LOW_WATER)
+
+        if count < threshold_high:
+            return 0
+
+        total_evicted = 0
+        while count >= threshold_low:
+            if total_evicted > 0 and count < threshold_low:
+                break
+            deleted = await evict_batch(effective_batch)
+            if deleted == 0:
+                break
+            total_evicted += deleted
+            count -= deleted
+            if count < threshold_high and total_evicted > 0:
+                break
+            if count < threshold_low:
+                break
+
+        if total_evicted > 0:
+            entry = {
+                "timestamp": datetime.now(timezone.utc).isoformat(),
+                "count": total_evicted,
+                "remaining": count,
+            }
+            EVICTION_LOG.append(entry)
+            if len(EVICTION_LOG) > 1000:
+                EVICTION_LOG.pop(0)
+            log.info(f"Evicted {total_evicted} vectors ({count} remaining)")
+
+        return total_evicted
+
+
+async def get_rag_operational_stats() -> dict:
+    stats = await get_collection_stats()
+    now = datetime.now(timezone.utc)
+    cutoff_1m = now - timedelta(minutes=1)
+    cutoff_5m = now - timedelta(minutes=5)
+    cutoff_30m = now - timedelta(minutes=30)
+
+    eviction_1m = sum(
+        e["count"] for e in EVICTION_LOG
+        if datetime.fromisoformat(e["timestamp"]) > cutoff_1m
+    )
+    eviction_5m = sum(
+        e["count"] for e in EVICTION_LOG
+        if datetime.fromisoformat(e["timestamp"]) > cutoff_5m
+    )
+    eviction_30m = sum(
+        e["count"] for e in EVICTION_LOG
+        if datetime.fromisoformat(e["timestamp"]) > cutoff_30m
+    )
+
+    pinned_count = 0
+    avg_retrieval_count = 0.0
+    at_risk_count = 0
+
+    try:
+        async with httpx.AsyncClient() as client:
+            pinned_filter = {
+                "should": [
+                    {"match": {"key": "source", "value": src}}
+                    for src in RAG_PINNED_SOURCES
+                ]
+            }
+            pinned_resp = await client.post(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}/points/scroll",
+                json={"filter": pinned_filter, "limit": 10000, "with_payload": True, "with_vector": False},
+                timeout=10.0,
+            )
+            if pinned_resp.status_code == 200:
+                pinned_count = len(pinned_resp.json().get("result", {}).get("points", []))
+
+            nonpinned_filter = {
+                "must_not": [
+                    {"match": {"key": "source", "value": src}}
+                    for src in RAG_PINNED_SOURCES
+                ]
+            }
+            np_resp = await client.post(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}/points/scroll",
+                json={"filter": nonpinned_filter, "limit": 10000, "with_payload": True, "with_vector": False},
+                timeout=10.0,
+            )
+            if np_resp.status_code == 200:
+                points = np_resp.json().get("result", {}).get("points", [])
+                if points:
+                    retrievals = []
+                    scored = []
+                    for p in points:
+                        payload = p.get("payload", {})
+                        rc = payload.get("retrieval_count", 0) or 0
+                        retrievals.append(rc)
+                        date_str = payload.get("ingest_date") or payload.get("upload_date", "")
+                        if date_str:
+                            age_hours = (now - datetime.fromisoformat(date_str)).total_seconds() / 3600
+                        else:
+                            age_hours = 999999
+                        score = rc * RAG_ACCESS_WEIGHT + age_hours * RAG_AGE_WEIGHT
+                        last_accessed = payload.get("last_accessed", date_str)
+                        scored.append((score, last_accessed))
+
+                    avg_retrieval_count = round(sum(retrievals) / len(retrievals), 2)
+
+                    scored.sort(key=lambda x: (x[0], x[1]))
+                    at_risk_threshold = max(1, len(scored) // 10)
+                    at_risk_count = at_risk_threshold
+    except Exception as e:
+        log.warning(f"RAG operational stats scroll error: {e}")
+
+    stats.update({
+        "grace_hours": RAG_GRACE_HOURS,
+        "eviction_counts_last_1m": eviction_1m,
+        "eviction_counts_last_5m": eviction_5m,
+        "eviction_counts_last_30m": eviction_30m,
+        "pinned_count": pinned_count,
+        "avg_retrieval_count": avg_retrieval_count,
+        "at_risk_count": at_risk_count,
+    })
+    return stats

rag.py

@@ -1,26 +1,31 @@
 """
 JarvisChat - RAG pipeline: Qdrant vector search + system prompt assembly.
 """
+import asyncio
 import logging
 
 import httpx
 
+from eviction import _update_retrieval_count
 from db import get_db, get_setting, list_skills_with_state, format_active_skills_prompt
 from memory import search_memories
-from config import MAX_SKILL_PROMPT_CHARS
+from config import MAX_SKILL_PROMPT_CHARS, QDRANT_URL, RAG_COLLECTION
 
 log = logging.getLogger("jarvischat")
 
-QDRANT_URL = "http://192.168.50.108:6333"
 EMBED_URL = "http://192.168.50.210:11434"
 EMBED_MODEL = "mxbai-embed-large"
-RAG_COLLECTION = "jarvis_rag"
 RAG_SCORE_THRESHOLD = 0.25
 
+# Re-export eviction symbols for backward compatibility
+from eviction import (  # noqa: E402
+    maybe_evict, get_rag_operational_stats, EVICTION_LOG,
+    get_collection_count, get_collection_stats, evict_batch,
+)
+
 
 def chunk_text(text: str, chunk_size: int = 512, overlap: int = 128) -> list:
     words = text.split()
-    # Approximate token count as len(words) * 1.3
     target_words = int(chunk_size / 1.3)
     overlap_words = int(overlap / 1.3)
     if not words:
@@ -54,7 +59,13 @@ async def query_rag(query: str, limit: int = 3) -> list:
             )
             if search_resp.status_code != 200:
                 return []
-            return search_resp.json().get("result", [])
+            results = search_resp.json().get("result", [])
+            for r in results:
+                pid = r.get("id")
+                if pid:
+                    current = r.get("payload", {}).get("retrieval_count", 0) or 0
+                    asyncio.ensure_future(_update_retrieval_count(pid, current))
+            return results
     except Exception as e:
         log.warning(f"RAG query error: {e}")
         return []

routers/ingest.py

@@ -7,6 +7,7 @@ from fastapi import APIRouter, HTTPException, Request
 from fastapi.responses import JSONResponse
 
 from config import COMPLETIONS_API_KEY
+from eviction import maybe_evict
 from rag import chunk_text, QDRANT_URL, EMBED_URL, EMBED_MODEL, RAG_COLLECTION
 
 log = logging.getLogger("jarvischat")
@@ -61,4 +62,9 @@ async def ingest_content(request: Request):
             else:
                 log.warning(f"Ingest Qdrant upsert failed for chunk {i}: {upsert_resp.status_code}")
 
+    if ingested > 0:
+        evicted = await maybe_evict()
+        if evicted:
+            log.info(f"Evicted {evicted} vectors after ingest")
+
     return {"chunks_ingested": ingested, "source": source, "message": f"Ingested {ingested} chunks from {source}"}

routers/rag_admin.py

@@ -0,0 +1,60 @@
+"""JarvisChat routers — RAG corpus management admin endpoints."""
+import logging
+
+import httpx
+from fastapi import APIRouter, HTTPException, Request
+from fastapi.responses import JSONResponse
+
+from eviction import get_rag_operational_stats, EVICTION_LOG
+from rag import QDRANT_URL, RAG_COLLECTION
+
+log = logging.getLogger("jarvischat")
+router = APIRouter()
+
+
+@router.get("/api/rag/stats")
+async def rag_stats(request: Request):
+    if getattr(request.state, "session_role", "none") != "admin":
+        raise HTTPException(status_code=403, detail="Admin PIN required for this action")
+    stats = await get_rag_operational_stats()
+    stats["eviction_log_size"] = len(EVICTION_LOG)
+    return stats
+
+
+@router.post("/api/rag/flush")
+async def rag_flush(request: Request):
+    try:
+        async with httpx.AsyncClient() as client:
+            scroll_resp = await client.post(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}/points/scroll",
+                json={"limit": 10000, "with_payload": False, "with_vector": False},
+                timeout=30.0,
+            )
+            if scroll_resp.status_code != 200:
+                return JSONResponse(status_code=502, content={"detail": f"Qdrant scroll failed: {scroll_resp.status_code}"})
+
+            all_points = scroll_resp.json().get("result", {}).get("points", [])
+            point_ids = [p["id"] for p in all_points]
+
+            if not point_ids:
+                return {"deleted_count": 0, "collection": RAG_COLLECTION, "status": "flushed"}
+
+            delete_resp = await client.post(
+                f"{QDRANT_URL}/collections/{RAG_COLLECTION}/points/delete",
+                json={"points": point_ids},
+                timeout=30.0,
+            )
+            if delete_resp.status_code not in (200, 201):
+                return JSONResponse(status_code=502, content={"detail": f"Qdrant delete failed: {delete_resp.status_code}"})
+
+            EVICTION_LOG.clear()
+            log.warning(f"RAG collection '{RAG_COLLECTION}' flushed ({len(point_ids)} points deleted)")
+
+            return {
+                "deleted_count": len(point_ids),
+                "collection": RAG_COLLECTION,
+                "status": "flushed",
+            }
+    except Exception as e:
+        log.warning(f"RAG flush error: {e}")
+        return JSONResponse(status_code=502, content={"detail": f"RAG flush error: {e}"})

routers/upload.py

@@ -10,6 +10,7 @@ from fastapi.responses import JSONResponse
 
 from config import UPLOAD_DIR, MAX_UPLOAD_BYTES, SUPPORTED_UPLOAD_TYPES, UPLOAD_CONTEXT_EXPIRY_HOURS
 from db import get_db, insert_upload_context, list_upload_context_by_conversation, delete_upload_context_by_id
+from eviction import maybe_evict
 from rag import chunk_text, QDRANT_URL, EMBED_URL, EMBED_MODEL, RAG_COLLECTION
 
 log = logging.getLogger("jarvischat")
@@ -87,6 +88,10 @@ async def upload_file(
                 else:
                     log.warning(f"Qdrant upsert failed for chunk {i}: {upsert_resp.status_code}")
         result["chunks_ingested"] = ingested
+        if ingested > 0:
+            evicted = await maybe_evict()
+            if evicted:
+                log.info(f"Evicted {evicted} vectors after upload")
 
     if mode in ("context", "both"):
         expires = (datetime.now(timezone.utc) + timedelta(hours=UPLOAD_CONTEXT_EXPIRY_HOURS)).isoformat()

tests/test_rag_management.py

@@ -0,0 +1,409 @@
+import asyncio
+import os
+from datetime import datetime, timezone, timedelta
+from pathlib import Path
+
+import httpx
+from fastapi.testclient import TestClient
+
+import app
+import config
+import db
+import rag
+from security import SESSIONS, PIN_ATTEMPTS, RATE_EVENTS
+
+
+def make_client(tmp_path: Path) -> TestClient:
+    os.environ["JARVISCHAT_ADMIN_PIN"] = "1234"
+    db.DB_PATH = tmp_path / "jarvischat-rag-mgmt.db"
+    SESSIONS.clear()
+    PIN_ATTEMPTS.clear()
+    RATE_EVENTS.clear()
+    db.init_db()
+    return TestClient(app.app, raise_server_exceptions=False)
+
+
+def _admin_headers(client: TestClient) -> dict:
+    login = client.post("/api/auth/login", json={"pin": "1234"}, headers={"Origin": "http://testserver"})
+    sid = login.json()["session_id"]
+    return {"X-Session-ID": sid, "Origin": "http://testserver"}
+
+
+def _guest_headers(client: TestClient) -> dict:
+    sid = client.post("/api/auth/guest", headers={"Origin": "http://testserver"}).json()["session_id"]
+    return {"X-Session-ID": sid, "Origin": "http://testserver"}
+
+
+class FakeResponse:
+    def __init__(self, status, json_data=None):
+        self.status_code = status
+        self._json = json_data or {}
+
+    def json(self):
+        return self._json
+
+
+class FakeAsyncClient:
+    def __init__(self, *a, **kw):
+        pass
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *a):
+        pass
+
+    async def get(self, url, **kw):
+        if "/collections/jarvis_rag" in url:
+            return FakeResponse(200, {"result": {"vectors_count": 123}})
+        return FakeResponse(200)
+
+    async def post(self, url, **kw):
+        if "/points/scroll" in url:
+            return FakeResponse(200, {"result": {"points": []}})
+        if "/points/delete" in url:
+            return FakeResponse(200)
+        return FakeResponse(200)
+
+    async def put(self, url, **kw):
+        return FakeResponse(200)
+
+
+def _old_ts(hours_ago: float = 24) -> str:
+    return (datetime.now(timezone.utc) - timedelta(hours=hours_ago)).isoformat()
+
+
+def _young_ts(hours_ago: float = 0.1) -> str:
+    return (datetime.now(timezone.utc) - timedelta(hours=hours_ago)).isoformat()
+
+
+# ---------- get_collection_count ----------
+
+def test_get_collection_count(monkeypatch):
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: FakeAsyncClient())
+    count = asyncio.run(rag.get_collection_count())
+    assert count == 123
+
+
+# ---------- get_collection_stats ----------
+
+def test_get_collection_stats_shape(monkeypatch):
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: FakeAsyncClient())
+    stats = asyncio.run(rag.get_collection_stats())
+    assert stats["vector_count"] == 123
+    assert stats["max_vectors"] == 50000
+    assert stats["high_water_mark"] == 40000
+    assert stats["low_water_mark"] == 10000
+    assert stats["high_water_pct"] == 80
+    assert stats["low_water_pct"] == 20
+    assert 0 < stats["percent_full"] < 1
+    assert "upload" in stats["pinned_sources"]
+    assert "profile" in stats["pinned_sources"]
+
+
+# ---------- evict_batch ----------
+
+def test_evict_batch_excludes_pinned_sources(monkeypatch):
+    """Pinned sources ('upload', 'profile') should be in the must_not scroll filter."""
+    old = _old_ts(48)
+    # Only non-pinned points are returned (real Qdrant would honour must_not filter)
+    scroll_points = [
+        {"id": "old-data", "payload": {"source": "terminal", "ingest_date": old, "retrieval_count": 0}},
+    ]
+
+    class ScrollClient(FakeAsyncClient):
+        async def post(self, url, **kw):
+            if "/points/scroll" in url:
+                must_not = kw.get("json", {}).get("filter", {}).get("must_not", [])
+                pinned_values = [m["match"]["value"] for m in must_not]
+                assert "upload" in pinned_values
+                assert "profile" in pinned_values
+                return FakeResponse(200, {"result": {"points": scroll_points}})
+            if "/points/delete" in url:
+                return FakeResponse(200)
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: ScrollClient())
+    deleted = asyncio.run(rag.evict_batch(10))
+    assert deleted == 1
+
+
+def test_evict_batch_respects_grace_period(monkeypatch):
+    """Vectors younger than RAG_GRACE_HOURS should be skipped."""
+    old = _old_ts(48)
+    young = _young_ts(0.1)
+    scroll_points = [
+        {"id": "mature", "payload": {"source": "terminal", "ingest_date": old, "retrieval_count": 0}},
+        {"id": "newborn", "payload": {"source": "terminal", "ingest_date": young, "retrieval_count": 0}},
+    ]
+
+    class GraceClient(FakeAsyncClient):
+        async def post(self, url, **kw):
+            if "/points/scroll" in url:
+                return FakeResponse(200, {"result": {"points": scroll_points}})
+            if "/points/delete" in url:
+                deleted = kw.get("json", {}).get("points", [])
+                assert "newborn" not in deleted
+                return FakeResponse(200)
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: GraceClient())
+    deleted = asyncio.run(rag.evict_batch(10))
+    assert deleted == 1
+
+
+def test_evict_batch_respects_batch_size(monkeypatch):
+    """Only up to batch_size vectors should be deleted per call."""
+    old = _old_ts(48)
+    points = [{"id": f"p{i}", "payload": {"source": "terminal", "ingest_date": old, "retrieval_count": 0}} for i in range(50)]
+
+    class BatchClient(FakeAsyncClient):
+        async def post(self, url, **kw):
+            if "/points/delete" in url:
+                deleted = kw.get("json", {}).get("points", [])
+                assert len(deleted) == 10
+                return FakeResponse(200)
+            if "/points/scroll" in url:
+                return FakeResponse(200, {"result": {"points": points}})
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: BatchClient())
+    deleted = asyncio.run(rag.evict_batch(10))
+    assert deleted == 10
+
+
+def test_evict_batch_all_pinned_returns_zero(monkeypatch):
+    """If scroll returns nothing (all points filtered by must_not), evict_batch returns 0."""
+    class EmptyClient(FakeAsyncClient):
+        async def post(self, url, **kw):
+            if "/points/scroll" in url:
+                return FakeResponse(200, {"result": {"points": []}})
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: EmptyClient())
+    deleted = asyncio.run(rag.evict_batch(10))
+    assert deleted == 0
+
+
+def test_evict_batch_scores_lowest_first(monkeypatch):
+    """Vectors with lower scores should be evicted first."""
+    old = _old_ts(48)
+    points = [
+        {"id": "high-score", "payload": {"source": "terminal", "ingest_date": old, "retrieval_count": 100}},
+        {"id": "low-score", "payload": {"source": "terminal", "ingest_date": old, "retrieval_count": 0}},
+        {"id": "mid-score", "payload": {"source": "terminal", "ingest_date": old, "retrieval_count": 50}},
+    ]
+
+    class ScoreClient(FakeAsyncClient):
+        async def post(self, url, **kw):
+            if "/points/delete" in url:
+                deleted = kw.get("json", {}).get("points", [])
+                assert "low-score" in deleted
+                assert "high-score" not in deleted
+                assert "mid-score" not in deleted
+                return FakeResponse(200)
+            if "/points/scroll" in url:
+                return FakeResponse(200, {"result": {"points": points}})
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: ScoreClient())
+    deleted = asyncio.run(rag.evict_batch(1))
+    assert deleted == 1
+
+
+# ---------- maybe_evict ----------
+
+def test_maybe_evict_below_high_water(monkeypatch):
+    """When count is below high-water mark, eviction should not fire."""
+    class LowCountClient(FakeAsyncClient):
+        async def get(self, url, **kw):
+            # 30000 < 40000 high water
+            return FakeResponse(200, {"result": {"vectors_count": 30000}})
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: LowCountClient())
+    rag.EVICTION_LOG.clear()
+    evicted = asyncio.run(rag.maybe_evict())
+    assert evicted == 0
+    assert len(rag.EVICTION_LOG) == 0
+
+
+def test_maybe_evict_at_high_water(monkeypatch):
+    """When count reaches high-water mark, eviction should fire."""
+    class HighCountClient(FakeAsyncClient):
+        def __init__(self, *a, **kw):
+            super().__init__()
+            self.call_count = 0
+
+        async def get(self, url, **kw):
+            return FakeResponse(200, {"result": {"vectors_count": 45000}})
+
+        async def post(self, url, **kw):
+            if "/points/scroll" in url:
+                old = _old_ts(48)
+                points = [{"id": f"evict-me-{i}", "payload": {"source": "terminal", "ingest_date": old, "retrieval_count": 0}} for i in range(100)]
+                return FakeResponse(200, {"result": {"points": points}})
+            if "/points/delete" in url:
+                return FakeResponse(200)
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: HighCountClient())
+    rag.EVICTION_LOG.clear()
+    evicted = asyncio.run(rag.maybe_evict())
+    assert evicted > 0
+    assert len(rag.EVICTION_LOG) == 1
+    entry = rag.EVICTION_LOG[0]
+    assert "timestamp" in entry
+    assert entry["count"] > 0
+
+
+def test_maybe_evict_zero_config_disabled(monkeypatch):
+    """RAG_MAX_VECTORS <= 0 should disable eviction."""
+    orig = config.RAG_MAX_VECTORS
+    try:
+        config.RAG_MAX_VECTORS = 0
+        evicted = asyncio.run(rag.maybe_evict())
+        assert evicted == 0
+    finally:
+        config.RAG_MAX_VECTORS = orig
+
+
+def test_maybe_evict_all_pinned_breaks(monkeypatch):
+    """Above high water but only pinned points exist → eviction breaks with 0 deleted."""
+    class AllPinnedClient(FakeAsyncClient):
+        async def get(self, url, **kw):
+            return FakeResponse(200, {"result": {"vectors_count": 45000}})
+        async def post(self, url, **kw):
+            if "/points/scroll" in url:
+                return FakeResponse(200, {"result": {"points": []}})
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: AllPinnedClient())
+    rag.EVICTION_LOG.clear()
+    evicted = asyncio.run(rag.maybe_evict())
+    assert evicted == 0
+    assert len(rag.EVICTION_LOG) == 0
+
+
+# ---------- get_rag_operational_stats ----------
+
+def test_rag_operational_stats_shape(monkeypatch):
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: FakeAsyncClient())
+    rag.EVICTION_LOG.clear()
+    rag.EVICTION_LOG.append({
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "count": 500,
+        "remaining": 40000,
+    })
+    stats = asyncio.run(rag.get_rag_operational_stats())
+    assert stats["vector_count"] == 123
+    assert stats["grace_hours"] == 1
+    assert "eviction_counts_last_1m" in stats
+    assert "eviction_counts_last_5m" in stats
+    assert "eviction_counts_last_30m" in stats
+    assert stats["eviction_counts_last_1m"] == 500
+
+
+# ---------- GET /api/rag/stats ----------
+
+def test_rag_stats_endpoint(tmp_path, monkeypatch):
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: FakeAsyncClient())
+    with make_client(tmp_path) as client:
+        resp = client.get("/api/rag/stats", headers=_admin_headers(client))
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["vector_count"] == 123
+        assert data["max_vectors"] == 50000
+        assert "high_water_mark" in data
+        assert "low_water_mark" in data
+        assert data["high_water_pct"] == 80
+        assert data["low_water_pct"] == 20
+        assert "percent_full" in data
+        assert data["pinned_sources"] == ["upload", "profile"]
+        assert data["grace_hours"] == 1
+        assert "eviction_counts_last_1m" in data
+        assert "eviction_counts_last_5m" in data
+        assert "eviction_counts_last_30m" in data
+        assert "pinned_count" in data
+        assert "avg_retrieval_count" in data
+        assert "at_risk_count" in data
+        assert "eviction_log_size" in data
+
+
+def test_rag_stats_requires_admin(tmp_path, monkeypatch):
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: FakeAsyncClient())
+    with make_client(tmp_path) as client:
+        resp = client.get("/api/rag/stats", headers=_guest_headers(client))
+        assert resp.status_code == 403
+
+
+# ---------- POST /api/rag/flush ----------
+
+def test_rag_flush_endpoint(tmp_path, monkeypatch):
+    class FlushClient(FakeAsyncClient):
+        async def post(self, url, **kw):
+            if "/points/scroll" in url:
+                return FakeResponse(200, {"result": {"points": [{"id": "a"}, {"id": "b"}]}})
+            if "/points/delete" in url:
+                return FakeResponse(200)
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: FlushClient())
+    with make_client(tmp_path) as client:
+        resp = client.post("/api/rag/flush", headers=_admin_headers(client))
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["status"] == "flushed"
+        assert data["deleted_count"] == 2
+        assert data["collection"] == rag.RAG_COLLECTION
+
+
+def test_rag_flush_requires_admin(tmp_path, monkeypatch):
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: FakeAsyncClient())
+    with make_client(tmp_path) as client:
+        resp = client.post("/api/rag/flush", headers=_guest_headers(client))
+        assert resp.status_code == 403
+
+
+def test_rag_flush_empty_collection(tmp_path, monkeypatch):
+    class EmptyClient(FakeAsyncClient):
+        async def post(self, url, **kw):
+            if "/points/scroll" in url:
+                return FakeResponse(200, {"result": {"points": []}})
+            return FakeResponse(200)
+
+    monkeypatch.setattr(httpx, "AsyncClient", lambda *a, **kw: EmptyClient())
+    with make_client(tmp_path) as client:
+        resp = client.post("/api/rag/flush", headers=_admin_headers(client))
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["deleted_count"] == 0
+        assert data["status"] == "flushed"
+
+
+# ---------- Race lock ----------
+
+def test_eviction_lock_prevents_concurrent_eviction(monkeypatch):
+    """Concurrent calls to maybe_evict should queue; only one evicts."""
+    call_order = []
+
+    async def slow_get_collection_count():
+        call_order.append("count")
+        return 45000
+
+    async def slow_evict_batch(bs):
+        call_order.append("evict")
+        await asyncio.sleep(0.05)
+        return 500
+
+    monkeypatch.setattr(rag, "get_collection_count", slow_get_collection_count)
+    monkeypatch.setattr(rag, "evict_batch", slow_evict_batch)
+    rag.EVICTION_LOG.clear()
+
+    async def run_concurrent():
+        r1, r2 = await asyncio.gather(rag.maybe_evict(), rag.maybe_evict())
+        return r1, r2
+
+    r1, r2 = asyncio.run(run_concurrent())
+    # First call evicted, second found count already below high water or lock serialized
+    assert r1 >= 0
+    assert r2 >= 0