# Project export: Narcore This document was generated by HackStack to give an AI agent context about a hackathon project. Sections are labeled with their provenance; content marked as truncated was cut to keep this document small. ## Project metadata - Hackathon: UC Berkeley AI Hackathon 2026 - Tagline: An AI undercover agent to detect and intercept online drug advertising. - Devpost: https://devpost.com/software/narcore - GitHub: https://github.com/cibarbia05/narcore - Video: https://www.youtube.com/embed/r2Dl7fKbcsU?enablejsapi=1&hl=en_US&rel=0&start=&version=3&wmode=transparent - Result: winner (Best Use of Browserbase) - Team: 3 GitHub contributor(s) — cibarbia05 (29 commits), Claude Opus 4.8 (1M context) (13 commits), durpdur (8 commits) ## Devpost submission (written by the team) No Devpost description available. ## README (from the GitHub repository) # narcore A dark, focused web foundation for high-stakes government tooling. ## Stack - **[Next.js](https://nextjs.org) (App Router)** + **React 19** + **TypeScript** - **[Tailwind CSS v4](https://tailwindcss.com)** — CSS-first config via `@theme` (no `tailwind.config.js`) - **[shadcn/ui](https://ui.shadcn.com)** (Base UI primitives, Lucide icons) - **[next-themes](https://github.com/pacocoursey/next-themes)** — dark-first theming ## Getting started ```bash pnpm install pnpm dev # http://localhost:3000 ``` Other scripts: ```bash pnpm build # production build (type-check + lint + compile) pnpm start # serve the production build pnpm lint # ESLint ``` > Uses **pnpm**. Install it with `npm i -g pnpm` (or via Corepack) if needed. ## Project layout ``` src/ app/ layout.tsx # fonts, dark-default ThemeProvider, metadata globals.css # design tokens — single source of truth (dark + azure) page.tsx # starter page components/ ui/ # shadcn/ui primitives (e.g. button) theme-provider.tsx logo.tsx # brand mark lib/ utils.ts # cn() class merge helper components.json # shadcn/ui config brand.md # brand & UI guidelines ``` # NARCORE — Technical Architecture A field guide to how the system actually works. Focused on the core engine: the parallel scraper fleets, the undercover outreach operative, and the infrastructure underneath them (Browserbase, Redis, embeddings, and the LLMs). --- ## 1. The Big Picture NARCORE is a closed-loop system that does three things and feeds each one back into the next: 1. **Detect** — parallel browser agents scrape Instagram hashtags, and every post is risk-scored using semantic (vector) similarity against a corpus of known drug-dealing language. 2. **Engage** — a human operator launches an undercover "operative" against a flagged seller. An LLM-driven agent negotiates over Instagram DMs toward two objectives: confirm the deal and confirm a meeting location. 3. **Learn** — every confirmed engagement teaches the system. New coded slang is extracted from the seller's own words and added to the detection corpus (R1), and the successful tactics are stored as long-term memory to prime the next operative (R2). ``` ┌─────────────┐ flagged ┌──────────────┐ confirmed ┌────────────┐ │ DETECT │ ───────────────▶│ ENGAGE │ ───────────────▶ │ LEARN │ │ scraper │ lead │ operative │ deal+location │ R1 corpus │ │ fleet (×N) │ │ (LLM + DM) │ │ R2 memory │ └─────────────┘ └──────────────┘ └────────────┘ ▲ │ └───────────────── grown corpus re-flags more posts ◀──────────────┘ ``` Everything is coordinated through a single **Redis** instance, which acts as the vector database, the state store, the job registry, and the event bus all at once. --- ## 2. Parallel Scraper Fleets **What it is:** N independent browser agents (default 5, configurable 1–20) that each drive a real Instagram session in parallel, scrape a hashtag feed, and push every post into the scoring pipeline. **Key files:** `src/lib/agents/orchestrator.ts`, `src/lib/agents/ig-agent.ts`, `src/lib/agents/run-store.ts` ### How a fleet launches 1. `POST /api/agents/run` → `startRun()` resolves the agent count and assigns each agent a distinct Instagram hashtag target. 2. All N Browserbase sessions are provisioned **up front and in parallel** (`Promise.all` over `provisionAgent()`), so the live-view URLs exist before any scraping begins — the UI can render the video grid immediately. 3. Each agent's loop (`runAgentLoop`) is fired **un-awaited** (`void`). The HTTP request returns instantly; the agents keep running in the background. 4. The UI **polls** Redis for progress rather than holding an open connection. ### Why there are no race conditions Every agent owns its own Redis hash: `run:{id}:agent:{idx}`. An agent only ever patches its own key, so five siblings writing concurrently never collide — no locks needed. The run is marked `done` only when *all* agents reach a terminal state (`done`, `blocked`, `error`, or `stopped`). ### Budgets (per agent) | Setting | Default | Meaning | |---|---|---| | `IG_AGENT_TIMEOUT_MS` | 120s | Wall-clock budget per agent | | `IG_MAX_POSTS_PER_AGENT` | 8 | Posts captured before stopping | ### Cancellation `stopRun()` immediately marks the run stopped, aborts every agent's `AbortController`, and releases all Browserbase sessions in parallel. The UI sees a clean terminal state right away. --- ## 3. The Undercover Operative (Outreach) **What it is:** A single LLM-driven agent that opens an Instagram DM with a flagged seller and negotiates, turn by turn, toward two confirmations. Unlike the fleet, operations run **one at a time per post** (deduped) and have a much longer budget because real conversations take time. **Key files:** `src/lib/agents/operation-orchestrator.ts`, `scraper/operative-agent.ts`, `src/lib/operative-brain.ts`, `src/lib/agents/operation-store.ts` ### Launch flow (`startOperation`) 1. **Preconditions:** `ANTHROPIC_API_KEY` must be set, at least one logged-in Browserbase context must exist, and the target handle must pass the **allowlist gate** (`OPERATIVE_ALLOWLIST_ENFORCED=true` — this must never be off in a live demo; it restricts targets to consented demo accounts). 2. **Dedup:** `op:by-post:{postId}` ensures the same post can't launch two operations at once. 3. A Browserbase session is created, an `Operation` record is written to Redis with status `opening`, and the operative loop fires un-awaited (same fire-and-forget + poll pattern as the fleet). ### The negotiation loop The operative tracks two objectives independently so the UI can show live state like **"Deal ✓ · Location ✗"** even mid-conversation. ``` open DM thread → brain composes opener → ┌─────────────────────────────┐ │ send message (verify it │ │ landed, up to 3 retries) │ │ wait for seller reply │ │ brain analyzes transcript │ │ patch deal/location state │ └──────────────┬──────────────┘ │ both confirmed? ── yes ──▶ send closer, mark "confirmed", trigger R1+R2 │ no rejected / stalled / max turns? ── yes ──▶ terminal │ no └──▶ loop ``` Opening a DM thread uses one of two strategies (`OPERATIVE_DM_OPEN_STRATEGY`): - **`ladder`** (default): a hand-rolled `observe → act` sequence (try the profile "Message" button, fall back to DM-inbox search). - **`agent`**: a self-healing DOM-mode Stagehand agent that figures out the steps itself (more robust, more expensive). ### Budgets (per operation) | Setting | Default | Meaning | |---|---|---| | `OPERATIVE_BUDGET_MS` | 25 min | Total wall-clock per negotiation | | `OPERATIVE_REPLY_WAIT_MS` | 4 min | Max wait for each seller reply | | `OPERATIVE_POLL_MS` | 15s | How often the thread is checked for new replies | | `OPERATIVE_MAX_TURNS` | 12 | Max back-and-forth exchanges | ### Operation states `opening → awaiting_reply → analyzing → negotiating` and then a terminal state: `confirmed` (both objectives met), `rejected`, `stalled` (ran out of budget/turns), `blocked` (login wall), `error`, or `stopped` (operator aborted). ### The Operative Brain (the LLM call) `negotiate()` in `src/lib/operative-brain.ts` makes **one Claude call per turn**. It is given the lead contex [README truncated for size] ## Detected evidence (automated analysis) Indexed codebase: 458 recognized source files, 1875 KB. - Anthropic (technology) — detected in the code - CSS (language) — detected in the code - JavaScript (language) — detected in the code - Next.js (technology) — detected in the code - React (technology) — detected in the code - Redis (technology) — detected in the code - Tailwind CSS (technology) — detected in the code - TypeScript (language) — detected in the code - Docker (technology) — claimed on Devpost, not found in the code - AI coding agent: Claude Code — evidence: config files committed to the repository; commit authorship or trailers - AI coding agent: Codex — evidence: config files committed to the repository ## Codebase structure (from repository index) ### Files (120 of 504) ``` .agents/skills/browser/EXAMPLES.md .agents/skills/browser/LICENSE.txt .agents/skills/browser/REFERENCE.md .agents/skills/browser/SKILL.md .agents/skills/python-patterns/SKILL.md .agents/skills/redis-connections/.cursor-plugin/plugin.json .agents/skills/redis-connections/references/blocking.md .agents/skills/redis-connections/references/client-cache.md .agents/skills/redis-connections/references/pipelining.md .agents/skills/redis-connections/references/pooling.md .agents/skills/redis-connections/references/timeouts.md .agents/skills/redis-connections/SKILL.md .agents/skills/redis-core/.cursor-plugin/plugin.json .agents/skills/redis-core/evals/core/baselines/aggregate-benchmark.json .agents/skills/redis-core/evals/core/baselines/aggregate-benchmark.md .agents/skills/redis-core/evals/core/baselines/baseline.json .agents/skills/redis-core/evals/core/baselines/model-matrix.json .agents/skills/redis-core/evals/core/baselines/README.md .agents/skills/redis-core/evals/core/evals.json .agents/skills/redis-core/evals/core/model-matrix.json .agents/skills/redis-core/references/choose-data-structure.md .agents/skills/redis-core/references/key-naming.md .agents/skills/redis-core/SKILL.md .agents/skills/redis-observability/.cursor-plugin/plugin.json .agents/skills/redis-observability/references/commands.md .agents/skills/redis-observability/references/metrics.md .agents/skills/redis-observability/SKILL.md .agents/skills/redis-semantic-cache/.cursor-plugin/plugin.json .agents/skills/redis-semantic-cache/references/best-practices.md .agents/skills/redis-semantic-cache/references/langcache-usage.md .agents/skills/redis-semantic-cache/SKILL.md .agents/skills/redis-vector-search/.cursor-plugin/plugin.json .agents/skills/redis-vector-search/references/algorithm-choice.md .agents/skills/redis-vector-search/references/hybrid-search.md .agents/skills/redis-vector-search/references/index-creation.md .agents/skills/redis-vector-search/references/rag-pattern.md .agents/skills/redis-vector-search/SKILL.md .agents/skills/security-review/SKILL.md .agents/skills/vercel-composition-patterns/AGENTS.md .agents/skills/vercel-composition-patterns/metadata.json .agents/skills/vercel-composition-patterns/README.md .agents/skills/vercel-composition-patterns/rules/_sections.md .agents/skills/vercel-composition-patterns/rules/_template.md .agents/skills/vercel-composition-patterns/rules/architecture-avoid-boolean-props.md .agents/skills/vercel-composition-patterns/rules/architecture-compound-components.md .agents/skills/vercel-composition-patterns/rules/patterns-children-over-render-props.md .agents/skills/vercel-composition-patterns/rules/patterns-explicit-variants.md .agents/skills/vercel-composition-patterns/rules/react19-no-forwardref.md .agents/skills/vercel-composition-patterns/rules/state-context-interface.md .agents/skills/vercel-composition-patterns/rules/state-decouple-implementation.md .agents/skills/vercel-composition-patterns/rules/state-lift-state.md .agents/skills/vercel-composition-patterns/SKILL.md .agents/skills/vercel-react-best-practices/AGENTS.md .agents/skills/vercel-react-best-practices/metadata.json .agents/skills/vercel-react-best-practices/README.md .agents/skills/vercel-react-best-practices/rules/_sections.md .agents/skills/vercel-react-best-practices/rules/_template.md .agents/skills/vercel-react-best-practices/rules/advanced-effect-event-deps.md .agents/skills/vercel-react-best-practices/rules/advanced-event-handler-refs.md .agents/skills/vercel-react-best-practices/rules/advanced-init-once.md .agents/skills/vercel-react-best-practices/rules/advanced-use-latest.md .agents/skills/vercel-react-best-practices/rules/async-api-routes.md .agents/skills/vercel-react-best-practices/rules/async-cheap-condition-before-await.md .agents/skills/vercel-react-best-practices/rules/async-defer-await.md .agents/skills/vercel-react-best-practices/rules/async-dependencies.md .agents/skills/vercel-react-best-practices/rules/async-parallel.md .agents/skills/vercel-react-best-practices/rules/async-suspense-boundaries.md .agents/skills/vercel-react-best-practices/rules/bundle-analyzable-paths.md .agents/skills/vercel-react-best-practices/rules/bundle-barrel-imports.md .agents/skills/vercel-react-best-practices/rules/bundle-conditional.md .agents/skills/vercel-react-best-practices/rules/bundle-defer-third-party.md .agents/skills/vercel-react-best-practices/rules/bundle-dynamic-imports.md .agents/skills/vercel-react-best-practices/rules/bundle-preload.md .agents/skills/vercel-react-best-practices/rules/client-event-listeners.md .agents/skills/vercel-react-best-practices/rules/client-localstorage-schema.md .agents/skills/vercel-react-best-practices/rules/client-passive-event-listeners.md .agents/skills/vercel-react-best-practices/rules/client-swr-dedup.md .agents/skills/vercel-react-best-practices/rules/js-batch-dom-css.md .agents/skills/vercel-react-best-practices/rules/js-cache-function-results.md .agents/skills/vercel-react-best-practices/rules/js-cache-property-access.md .agents/skills/vercel-react-best-practices/rules/js-cache-storage.md .agents/skills/vercel-react-best-practices/rules/js-combine-iterations.md .agents/skills/vercel-react-best-practices/rules/js-early-exit.md .agents/skills/vercel-react-best-practices/rules/js-flatmap-filter.md .agents/skills/vercel-react-best-practices/rules/js-hoist-regexp.md .agents/skills/vercel-react-best-practices/rules/js-index-maps.md .agents/skills/vercel-react-best-practices/rules/js-length-check-first.md .agents/skills/vercel-react-best-practices/rules/js-min-max-loop.md .agents/skills/vercel-react-best-practices/rules/js-request-idle-callback.md .agents/skills/vercel-react-best-practices/rules/js-set-map-lookups.md .agents/skills/vercel-react-best-practices/rules/js-tosorted-immutable.md .agents/skills/vercel-react-best-practices/rules/rendering-activity.md .agents/skills/vercel-react-best-practices/rules/rendering-animate-svg-wrapper.md .agents/skills/vercel-react-best-practices/rules/rendering-conditional-render.md .agents/skills/vercel-react-best-practices/rules/rendering-content-visibility.md .agents/skills/vercel-react-best-practices/rules/rendering-hoist-jsx.md .agents/skills/vercel-react-best-practices/rules/rendering-hydration-no-flicker.md .agents/skills/vercel-react-best-practices/rules/rendering-hydration-suppress-warning.md .agents/skills/vercel-react-best-practices/rules/rendering-resource-hints.md .agents/skills/vercel-react-best-practices/rules/rendering-script-defer-async.md .agents/skills/vercel-react-best-practices/rules/rendering-svg-precision.md .agents/skills/vercel-react-best-practices/rules/rendering-usetransition-loading.md .agents/skills/vercel-react-best-practices/rules/rerender-defer-reads.md .agents/skills/vercel-react-best-practices/rules/rerender-dependencies.md .agents/skills/vercel-react-best-practices/rules/rerender-derived-state-no-effect.md .agents/skills/vercel-react-best-practices/rules/rerender-derived-state.md .agents/skills/vercel-react-best-practices/rules/rerender-functional-setstate.md .agents/skills/vercel-react-best-practices/rules/rerender-lazy-state-init.md .agents/skills/vercel-react-best-practices/rules/rerender-memo-with-default-value.md .agents/skills/vercel-react-best-practices/rules/rerender-memo.md .agents/skills/vercel-react-best-practices/rules/rerender-move-effect-to-event.md .agents/skills/vercel-react-best-practices/rules/rerender-no-inline-components.md .agents/skills/vercel-react-best-practices/rules/rerender-simple-expression-in-memo.md .agents/skills/vercel-react-best-practices/rules/rerender-split-combined-hooks.md .agents/skills/vercel-react-best-practices/rules/rerender-transitions.md .agents/skills/vercel-react-best-practices/rules/rerender-use-deferred-value.md .agents/skills/vercel-react-best-practices/rules/rerender-use-ref-transient-values.md .agents/skills/vercel-react-best-practices/rules/server-after-nonblocking.md .agents/skills/vercel-react-best-practices/rules/server-auth-actions.md .agents/skills/vercel-react-best-practices/rules/server-cache-lru.md [384 more files omitted for size] ``` ### Dependencies - package.json: @anthropic-ai/sdk@^0.105.0, @base-ui/react@^1.6.0, @browserbasehq/sdk@^2.10.0, @browserbasehq/stagehand@^3.6.0, @tailwindcss/postcss@^4, @types/node@^20, @types/react@^19, @types/react-dom@^19, class-variance-authority@^0.7.1, clsx@^2.1.1, eslint@^9, eslint-config-next@16.2.9, lucide-react@^1.21.0, next@16.2.9, next-themes@^0.4.6, react@19.2.4, react-dom@19.2.4, redis@^6.0.0, shadcn@^4.11.0, sonner@^2.0.7, swr@^2.4.1, tailwind-merge@^3.6.0, tailwindcss@^4, tsx@^4.22.4, tw-animate-css@^1.4.0, typescript@^5, umap-js@^1.4.0, zod@^4.4.3 ### Recent commits (newest first) - Revise README for NARCORE architecture overview - Add architecture doc - Add shared types, API hooks, and nav for the new features - Add docs for the Redis and Browserbase deepenings - Wire operative to field-intel, agent memory, and self-healing DM-open - Add consistent Browserbase session identity with match badge - Add Redis Iris agent memory for cross-operation operative recall - Add field-intel loop so confirmed busts teach the detection corpus - Update home page with 3d background and new text design - Enhance vector space UI to use UMAP - Fix fleet not showing leads vs posts difference properly - Merge branch 'main' of https://github.com/cibarbia05/gov-dr-ai - Add outreach via browserbase - Merge pull request #4 from cibarbia05/risk-rating - Risk rating formula updated - Merge branch 'main' of https://github.com/cibarbia05/gov-dr-ai - Add handling of up to 20 live views - Merge pull request #3 from cibarbia05/formula-update - formula change - Add computer use multi-agent view ## Key source files (fetched from GitHub, selected and truncated for size) ### AGENTS.md ```markdown @../.codex/AGENTS.md ## Codex Single source of truth: the shared instructions above live in the in-repo `.codex/AGENTS.md` (imported relatively so the import resolves for every teammate who clones this repo, regardless of their home directory). Codex reads that same `.codex/AGENTS.md` directly, so both tools stay aligned. Edit `.codex/AGENTS.md`, not this file. When making **any** UI change, invoke these skills to ensure best practices: - `vercel-react-best-practices` - `vercel-composition-patterns` - `web-design-guidelines` ``` ### brand.md ```markdown # Brand & UI Guidelines The visual system for **narcore** — a dark, focused interface for high-stakes government tooling. The bar is *world-class production*, not demo. Restraint is the defining quality: a near-black ground, generous space, precise type, and a **single authoritative-azure accent** that earns attention because nothing else competes for it. > **Source of truth:** all color/spacing/radius/motion values live as tokens in > [`src/app/globals.css`](./src/app/globals.css). This document explains *intent and > usage* — it never restates raw values that could drift out of sync. When a token > and this doc disagree, the token wins; fix the doc. --- ## 1. Principles 1. **Clarity over decoration.** Every element earns its place. If it doesn't aid comprehension or action, remove it. 2. **Restraint with the accent.** Azure marks the *one* thing that matters in a view — the primary action, focus, or a live signal. Never use it for large fills or as a background wash. 3. **Technical confidence.** Monospace for system/metadata, a clean grotesque for prose. The product should feel engineered, exact, and calm under pressure. 4. **Trust by default.** High contrast, predictable interaction, no dark patterns. This is software the public relies on. --- ## 2. Color Dark-first. The light theme exists for completeness, but the brand surface is dark (``, set in [`layout.tsx`](./src/app/layout.tsx)). Colors are authored in **OKLCH** for perceptually-even steps and predictable contrast. > **Severity ≠ brand.** Risk bands use dedicated semantic colors — red (`--destructive`) > for *High*, amber (`--chart-2`) for *Elevated*, neutral for *Low* — and are > intentionally independent of the azure accent. The brand color signals trust and > action, never danger. ### Semantic roles (token → usage) | Token | Role | | --- | --- | | `--background` / `--foreground` | Page ground (near-black, faintly warm) and primary text (soft off-white). | | `--card` / `--popover` | Raised surfaces, one step above the ground. | | `--primary` / `--primary-foreground` | **The azure accent** + the near-white text that sits on it. Primary buttons, focus, key emphasis. | | `--secondary` / `--accent` | Quiet neutral surfaces for hover and low-emphasis controls. | | `--muted` / `--muted-foreground` | Muted fills and secondary/supporting text. | | `--border` / `--input` | Hairline separators and field outlines — low-contrast, translucent white. | | `--ring` | Focus ring — azure, matches `--primary`. | | `--destructive` | Errors, irreversible actions, and the **High** risk band. | | `--chart-1…5` | Data viz — an azure-anchored multi-hue ramp; `--chart-2` is the amber used for the **Elevated** risk band. | | `--sidebar-*` | App-shell surfaces (reserved for future navigation chrome). | ### The one-accent rule - ✅ Primary CTA, focus ring, a live/status dot, a single emphasized number. - ❌ Azure headings, azure body text, large azure panels, two compe [truncated — 3294 more characters] ``` ### package.json ``` { "name": "gov-dr-ai", "version": "0.1.0", "private": true, "scripts": { "dev": "next dev", "build": "next build", "start": "next start", "lint": "eslint", "typecheck": "tsc --noEmit", "seed": "tsx scripts/seed.ts", "seed:memories": "tsx scripts/seed-memories.ts", "post-mock": "tsx scripts/post-mock.ts", "scrape-demo": "tsx scripts/scrape-demo.ts", "ig:login": "tsx scripts/ig-login.ts", "ig:verify": "tsx scripts/ig-verify.ts" }, "dependencies": { "@anthropic-ai/sdk": "^0.105.0", "@base-ui/react": "^1.6.0", "@browserbasehq/sdk": "^2.10.0", "@browserbasehq/stagehand": "^3.6.0", "class-variance-authority": "^0.7.1", "clsx": "^2.1.1", "lucide-react": "^1.21.0", "next": "16.2.9", "next-themes": "^0.4.6", "react": "19.2.4", "react-dom": "19.2.4", "redis": "^6.0.0", "shadcn": "^4.11.0", "sonner": "^2.0.7", "swr": "^2.4.1", "tailwind-merge": "^3.6.0", "tw-animate-css": "^1.4.0", "umap-js": "^1.4.0", "zod": "^4.4.3" }, "devDependencies": { "@tailwindcss/postcss": "^4", "@types/node": "^20", "@types/react": "^19", "@types/react-dom": "^19", "eslint": "^9", "eslint-config-next": "16.2.9", "tailwindcss": "^4", "tsx": "^4.22.4", "typescript": "^5" } } ``` ### docker-compose.yml ```yaml # Narcore local infrastructure. # Phase 0 ships a working Redis Stack so WT-B can integration-test for real. # WT-C is the SOLE editor of this file thereafter — it fleshes out the embedding # sidecar (llama.cpp serving nomic-embed-text-v2-moe on an OpenAI-compatible # /v1/embeddings endpoint). The Next.js app runs via `pnpm dev` (not containerized). services: redis: image: redis/redis-stack:latest container_name: narcore-redis ports: - "6379:6379" # Redis - "8001:8001" # RedisInsight (http://localhost:8001) volumes: - redis-data:/data healthcheck: test: ["CMD", "redis-cli", "ping"] interval: 5s timeout: 3s retries: 10 # --- WT-C: embedding sidecar --- # llama.cpp serving nomic-embed-text-v2-moe on an OpenAI-compatible POST /v1/embeddings. # We use the prebuilt ggml-org image (no local Dockerfile): its ENTRYPOINT is # /app/llama-server, so `command:` supplies args only. See infra/embedding/README.md. embedding: image: ghcr.io/ggml-org/llama.cpp:server # CPU build (multi-arch). GPU: see README (:server-cuda) container_name: narcore-embedding ports: - "8080:8080" # OpenAI-compatible POST /v1/embeddings — EMBEDDING_API_URL=http://localhost:8080/v1/embeddings # Pin the ggml-org GGUF (carries the v2-MoE pooling-assert conversion fix, #13534/#13689) # at its only file, the Q8_0 quant (near-lossless — embedding similarity is precision- # sensitive). --hf-file pins the exact file deterministically. The OpenAI /v1/embeddings # endpoint requires a non-`none` pooling type, hence --pooling mean. The app applies the # search_document:/search_query: prefixes itself — we must NOT. command: > -m /models/nomic-embed-text-v2-moe-q8_0.gguf --embeddings --pooling mean --host 0.0.0.0 --port 8080 -c 2048 -b 2048 -ub 2048 volumes: # GGUF is downloaded on the HOST (the container cannot reach Hugging Face) and # bind-mounted read-only. Fetch it once: see infra/embedding/README.md. - ./infra/embedding/models:/models:ro healthcheck: test: ["CMD", "curl", "-f", "http://localhost:8080/health"] interval: 10s timeout: 5s retries: 12 start_period: 300s # don't count health failures while the model downloads on first boot restart: unless-stopped # --- R2: Redis Iris agent memory (cross-operation operative memory) --- # Shared env for the agent-memory api + worker. Embeddings route through the # null-stripping proxy to the SAME local nomic embedder the app uses (one embedding # family, one Redis); generation uses Claude. Index is pinned to 768 dims to match # nomic. See OPERATIVE.md for why the proxy exists. embedding-proxy: image: node:20-alpine container_name: narcore-embed-proxy depends_on: - embedding command: ["node", "/app/server.js"] environment: UPSTREAM_URL: http://embedding:8080 PORT: "8090" volumes: - ./infra/embedding-proxy/server.js:/app/server.js:ro restart: unless-stopped agent-memory: image: redislabs/agent-memory-server:latest container_name: narcore-memory depends_on: redis: condition: service_healthy embedding-proxy: condition: service_started ports: - "8000:8000" # REST API — AGENT_MEMORY_URL=http://localhost:8000 environment: &agent-memory-env REDIS_URL: redis://redis:6379 DISABLE_AUTH: "true" AUTH_MODE: disabled GENERATION_MODEL: anthropic/claude-sonnet-4-6 FAST_MODEL: anthropic/claude-haiku-4-5 ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-} # Embeddings → null-stripping proxy → nomic (text-embedding-3-small is just a # name LiteLLM accepts; OPENAI_API_BASE points it at our local embedder). OPENAI_API_BASE: http://embedding-proxy:8090/v1 OPENAI_API_KEY: sk-noop EMBEDDING_MODEL: text-embedding-3-small REDISVL_VECTOR_DIMENSIONS: "768" # Explicit memories carry their own topics/entities — skip the heavy auto-extractors. ENABLE_NER: "false" ENABLE_TOPIC_EXTRACTION: "false" ENABLE_DISCRETE_MEMORY_EXTRACTION: "false" restart: unless-stopped # The Docket worker indexes long-term memories. It MUST run or memories are never # searchable (create returns ok but search stays empty). agent-memory-worker: image: redislabs/agent-memory-server:latest container_name: narcore-memory-worker depends_on: - agent-memory command: ["agent-memory", "task-worker"] environment: *agent-memory-env # The worker serves no HTTP port, so the image's API healthcheck would mark it # "unhealthy" even though it's indexing fine — disable it to avoid confusion. healthcheck: disable: true restart: unless-stopped volumes: redis-data: ``` ### src/app/layout.tsx ```typescript import type { Metadata } from "next"; import { Geist, Geist_Mono } from "next/font/google"; import "./globals.css"; import { ThemeProvider } from "@/components/theme-provider"; const fontSans = Geist({ variable: "--font-sans", subsets: ["latin"], }); const fontMono = Geist_Mono({ variable: "--font-mono", subsets: ["latin"], }); export const metadata: Metadata = { title: { default: "Narcore", template: "%s · Narcore", }, description: "Build something the public can trust.", }; export default function RootLayout({ children, }: Readonly<{ children: React.ReactNode; }>) { return ( {children} ); } ``` ### infra/embedding-proxy/server.js ```javascript // Tiny OpenAI-embeddings sanitizing proxy (zero deps). // // Why this exists: the Redis Agent Memory Server embeds via LiteLLM, which sends // `"encoding_format": null` on the OpenAI /v1/embeddings call. The llama.cpp server // that serves nomic-embed-text rejects a null where it expects a string // ("[json.exception.type_error.302] type must be string, but is null"). This proxy // sits between them and strips null-valued top-level fields from the JSON body, so // the memory server can use the same local nomic embedder the app uses — one // embedding family, one box, no external embedding key. // // Standalone CommonJS Node script run in a node:20-alpine container (`node // server.js`) — not part of the Next/TS app, hence require() over import. /* eslint-disable @typescript-eslint/no-require-imports */ const http = require("node:http"); const UPSTREAM = process.env.UPSTREAM_URL || "http://embedding:8080"; const PORT = Number(process.env.PORT || 8090); const upstream = new URL(UPSTREAM); function stripNulls(value) { if (Array.isArray(value)) return value.map(stripNulls); if (value && typeof value === "object") { for (const key of Object.keys(value)) { if (value[key] === null) delete value[key]; else value[key] = stripNulls(value[key]); } } return value; } const server = http.createServer((req, res) => { const chunks = []; req.on("data", (c) => chunks.push(c)); req.on("end", () => { let body = Buffer.concat(chunks); const contentType = req.headers["content-type"] || ""; if (contentType.includes("application/json") && body.length) { try { body = Buffer.from(JSON.stringify(stripNulls(JSON.parse(body.toString("utf8"))))); } catch { /* not JSON we can parse — forward unchanged */ } } const options = { hostname: upstream.hostname, port: upstream.port || 80, path: req.url, method: req.method, headers: { ...req.headers, host: upstream.host, "content-length": Buffer.byteLength(body) }, }; const proxied = http.request(options, (pres) => { res.writeHead(pres.statusCode || 502, pres.headers); pres.pipe(res); }); proxied.on("error", (err) => { res.writeHead(502, { "content-type": "application/json" }); res.end(JSON.stringify({ error: { message: String(err), type: "proxy_error" } })); }); proxied.end(body); }); }); server.listen(PORT, () => { console.log(`[embed-proxy] listening on :${PORT} -> ${UPSTREAM} (stripping null JSON fields)`); }); ``` ### src/app/page.tsx ```typescript import { ArrowRight } from "lucide-react"; import Link from "next/link"; import { Button } from "@/components/ui/button"; import { CodeSphereField } from "@/components/landing/code-sphere-field"; import { TypingHeadline } from "@/components/landing/typing-headline"; import { Logo } from "@/components/logo"; import { TopNav } from "@/components/top-nav"; export default function Home() { return (
{/* Restrained background: rotating code-spheres, faint azure glow + hairline grid. */}
{/* Ambient "depth of the corpus": code-glyph spheres rotating in 3D. */}
Law-enforcement AI for illicit-drug trafficking
narcore

Narcore scans social platforms for drug ads, sends an autonomous undercover operative to confirm the deal and meeting, and exports a court-ready case report.

); } ``` ### src/app/memory/page.tsx ```typescript import type { Metadata } from "next"; import { MemoryClient } from "@/components/memory/memory-client"; import { TopNav } from "@/components/top-nav"; export const metadata: Metadata = { title: "Operative Memory", description: "Redis Iris agent memory — what the operative has learned across operations.", }; export default function MemoryPage() { return ( <> ); } ``` ### src/app/semantic-drift/page.tsx ```typescript import type { Metadata } from "next"; import { SemanticDriftClient } from "@/components/semantic-drift/semantic-drift-client"; import { TopNav } from "@/components/top-nav"; export const metadata: Metadata = { title: "Semantic Drift", description: "Redis vector-space visualization for Narcore.", }; export default function SemanticDriftPage() { return ( <> ); } ``` ### src/app/(dashboard)/layout.tsx ```typescript // Route-group layout for the dashboard. Mounts the sonner here so toast // is scoped to the dashboard surface without touching the shared root layout. import { TopNav } from "@/components/top-nav"; import { Toaster } from "@/components/ui/sonner"; export default function DashboardLayout({ children, }: Readonly<{ children: React.ReactNode }>) { return ( <> {children} ); } ``` [435 more indexed source files omitted to keep this export small. The full file list is in the Codebase structure section above.]