adolf/README.md

# Adolf

Autonomous personal assistant with a multi-channel gateway. Three-tier model routing with GPU VRAM management.

## Architecture

```
┌─────────────────────────────────────────────────────┐
│                 CHANNEL ADAPTERS                    │
│                                                     │
│  [Telegram/Grammy]   [CLI]   [Voice — future]       │
│       ↕                ↕            ↕               │
│       └────────────────┴────────────┘               │
│                        ↕                            │
│          ┌─────────────────────────┐                │
│          │   GATEWAY  (agent.py)   │                │
│          │   FastAPI  :8000        │                │
│          │                         │                │
│          │  POST /message          │  ← all inbound │
│          │  POST /chat  (legacy)   │                │
│          │  GET  /stream/{id} SSE  │  ← token stream│
│          │  GET  /reply/{id}  SSE  │  ← legacy poll │
│          │  GET  /health           │                │
│          │                         │                │
│          │  channels.py registry   │                │
│          │  conversation buffers   │                │
│          └──────────┬──────────────┘                │
│                     ↓                               │
│          ┌──────────────────────┐                   │
│          │    AGENT CORE        │                   │
│          │  three-tier routing  │                   │
│          │  VRAM management     │                   │
│          └──────────────────────┘                   │
│                     ↓                               │
│          channels.deliver(session_id, channel, text)│
│               ↓                    ↓                │
│    telegram → POST grammy/send   cli → SSE queue    │
└─────────────────────────────────────────────────────┘
```

## Channel Adapters

| Channel | session_id | Inbound | Outbound |
|---------|-----------|---------|---------|
| Telegram | `tg-<chat_id>` | Grammy long-poll → POST /message | channels.py → POST grammy:3001/send |
| CLI | `cli-<user>` | POST /message directly | GET /stream/{id} SSE — Rich Live streaming |
| Voice | `voice-<device>` | (future) | (future) |

## Unified Message Flow

```
1. Channel adapter receives message
2. POST /message {text, session_id, channel, user_id}
3. 202 Accepted immediately
4. Background: run_agent_task(message, session_id, channel)
5. Parallel IO (asyncio.gather):
   a. _fetch_urls_from_message()       — Crawl4AI fetches any URLs in message
   b. _retrieve_memories()             — openmemory semantic search for context
   c. _fast_tool_runner.run_matching() — FastTools (weather, commute) if pattern matches
6. router.route() with enriched history (url_context + fast_context + memories)
   - fast tool match → force medium (real-time data, no point routing to light)
   - if URL content fetched and tier=light → upgrade to medium
7. Invoke agent for tier with url_context + memories in system prompt
8. Token streaming:
   - medium: astream() pushes per-token chunks to _stream_queues[session_id]; <think> blocks filtered in real time
   - light/complex: full reply pushed as single chunk after completion
   - _end_stream() sends [DONE] sentinel
9. channels.deliver(session_id, channel, reply_text) — Telegram callback
10. _store_memory() background task — stores turn in openmemory
11. GET /stream/{session_id} SSE clients receive chunks; CLI renders with Rich Live + final Markdown
```

## Tool Handling

Adolf uses LangChain's tool interface but only the complex agent actually invokes tools at runtime.

**Complex agent:** `web_search` and `fetch_url` are defined as `langchain_core.tools.Tool` objects and passed to `create_deep_agent()`. The deepagents library runs an agentic loop (LangGraph `create_react_agent` under the hood) that sends the tool schema to the model via OpenAI function-calling format and handles tool dispatch.

**Medium agent (default):** `_DirectModel` makes a single `model.ainvoke(messages)` call with no tool schema. Context (memories, fetched URL content) is injected via the system prompt instead. This is intentional — `qwen3:4b` behaves unreliably when a tool array is present.

**Memory tools (out-of-loop):** `add_memory` and `search_memory` are LangChain MCP tool objects (via `langchain_mcp_adapters`) but are excluded from both agents' tool lists. They are called directly — `await _memory_add_tool.ainvoke(...)` — outside the agent loop, before and after each turn.

## Three-Tier Model Routing

| Tier | Model | Agent | Trigger | Latency |
|------|-------|-------|---------|---------|
| Light | `qwen2.5:1.5b` (router answers directly) | — | Regex pre-match or 3-way embedding classifies "light" | ~2–4s |
| Medium | `qwen3:4b` (`DEEPAGENTS_MODEL`) | `_DirectModel` — single LLM call, no tools | Default; also forced when message contains URLs | ~10–20s |
| Complex | `deepseek/deepseek-r1:free` via LiteLLM (`DEEPAGENTS_COMPLEX_MODEL`) | `create_deep_agent` — agentic loop with tools | Auto-classified by embedding similarity | ~30–90s |

Routing is fully automatic via 3-way cosine similarity over pre-embedded utterance centroids (light / medium / complex). No prefix required. Use `adolf-deep` model name to force complex tier via API.

Complex tier is reached automatically for deep research queries — `исследуй`, `изучи все`, `напиши подробный`, etc. — via regex pre-classifier and embedding similarity. No prefix required. Use `adolf-deep` model name to force it via API.

## Fast Tools (`fast_tools.py`)

Pre-flight tools that run concurrently with URL fetch and memory retrieval before any LLM call. Each tool has two methods:
- `matches(message) → bool` — regex classifier; also used by `Router` to force medium tier
- `run(message) → str` — async fetch returning a context block injected into system prompt

`FastToolRunner` holds all tools. `any_matches()` is called by the Router at step 0a; `run_matching()` is called in the pre-flight `asyncio.gather` in `run_agent_task()`.

| Tool | Pattern | Source | Context returned |
|------|---------|--------|-----------------|
| `WeatherTool` | weather/forecast/temperature/snow/rain | SearXNG `"погода Балашиха сейчас"` | Current conditions in °C from Russian weather sites |
| `CommuteTool` | commute/traffic/arrival/пробки | `routecheck:8090/api/route` (Yandex Routing API) | Drive time with/without traffic, Balashikha→Moscow |

**To add a new fast tool:** subclass `FastTool` in `fast_tools.py`, implement `name`/`matches`/`run`, add an instance to `_fast_tool_runner` in `agent.py`.

## routecheck Service (`routecheck/`)

Local web service on port 8090. Exists because Yandex Routing API free tier requires a web UI that uses the API.

**Web UI** (`http://localhost:8090`): PIL-generated arithmetic captcha → lat/lon form → travel time result.

**Internal API**: `GET /api/route?from=lat,lon&to=lat,lon&token=ROUTECHECK_TOKEN` — bypasses captcha, used by `CommuteTool`. The `ROUTECHECK_TOKEN` shared secret is set in `.env` and passed to both `routecheck` and `deepagents` containers.

Yandex API calls are routed through the host HTTPS proxy (`host.docker.internal:56928`) since the container has no direct external internet access.

**Requires** `.env`: `YANDEX_ROUTING_KEY` (free from `developer.tech.yandex.ru`) + `ROUTECHECK_TOKEN`.

## Crawl4AI Integration

Crawl4AI runs as a Docker service (`crawl4ai:11235`) providing JS-rendered, bot-bypass page fetching.

**Pre-routing fetch (all tiers):**
- `_URL_RE` detects `https?://` URLs in any incoming message
- `_crawl4ai_fetch_async()` uses `httpx.AsyncClient` to POST `{urls: [...]}` to `/crawl`
- Up to 3 URLs fetched concurrently via `asyncio.gather`
- Fetched content (up to 3000 chars/URL) injected as a system context block into enriched history before routing and into medium/complex system prompts
- If fetch succeeds and router returns light → tier upgraded to medium

**Complex agent tools:**
- `web_search`: SearXNG query + Crawl4AI auto-fetch of top 2 result URLs → combined snippet + page text
- `fetch_url`: Crawl4AI single-URL fetch for any specific URL

## Memory Pipeline

openmemory runs as a FastMCP server (`openmemory:8765`) backed by mem0 + Qdrant + nomic-embed-text.

**Retrieval (before routing):** `_retrieve_memories()` calls `search_memory` MCP tool with the user message as query. Results (threshold ≥ 0.5) are prepended to enriched history so all tiers benefit.

**Storage (after reply):** `_store_memory()` runs as an asyncio background task, calling `add_memory` with `"User: ...\nAssistant: ..."`. The extraction LLM (`qwen2.5:1.5b` on GPU Ollama) pulls facts; dedup is handled by mem0's update prompt.

Memory tools (`add_memory`, `search_memory`, `get_all_memories`) are excluded from agent tool lists — memory management happens outside the agent loop.

## VRAM Management

GTX 1070 — 8 GB. Ollama must be restarted if CUDA init fails (model loads on CPU).

1. Flush explicitly before loading qwen3:8b (`keep_alive=0`)
2. Verify eviction via `/api/ps` poll (15s timeout) before proceeding
3. Fallback: timeout → run medium agent instead
4. Post-complex: flush 8b, pre-warm medium + router

## Session ID Convention

- Telegram: `tg-<chat_id>` (e.g. `tg-346967270`)
- CLI: `cli-<username>` (e.g. `cli-alvis`)

Conversation history is keyed by session_id (5-turn buffer).

## Files

```
adolf/
├── docker-compose.yml      Services: deepagents, openmemory, grammy, crawl4ai, routecheck, cli
├── Dockerfile              deepagents container (Python 3.12)
├── Dockerfile.cli          CLI container (python:3.12-slim + rich)
├── agent.py                FastAPI gateway, run_agent_task, Crawl4AI pre-fetch, fast tools, memory pipeline
├── fast_tools.py           FastTool base, FastToolRunner, WeatherTool, CommuteTool
├── channels.py             Channel registry + deliver() + pending_replies
├── router.py               Router class — regex + LLM tier classification, FastToolRunner integration
├── vram_manager.py         VRAMManager — flush/prewarm/poll Ollama VRAM
├── agent_factory.py        _DirectModel (medium) / create_deep_agent (complex)
├── cli.py                  Interactive CLI REPL — Rich Live streaming + Markdown render
├── wiki_research.py        Batch wiki research pipeline (uses /message + SSE)
├── benchmarks/
│   ├── run_benchmark.py    Routing accuracy benchmark — 120 queries across 3 tiers
│   ├── run_voice_benchmark.py  Voice path benchmark
│   ├── benchmark.json      Query dataset (gitignored)
│   └── results_latest.json Last run results (gitignored)
├── .env                    TELEGRAM_BOT_TOKEN, ROUTECHECK_TOKEN, YANDEX_ROUTING_KEY (not committed)
├── routecheck/
│   ├── app.py              FastAPI: image captcha + /api/route Yandex proxy
│   └── Dockerfile
├── tests/
│   ├── integration/        Standalone integration test scripts (common.py + test_*.py)
│   └── use_cases/          Claude Code skill markdown files — Claude acts as user + evaluator
├── openmemory/
│   ├── server.py           FastMCP + mem0: add_memory, search_memory, get_all_memories
│   └── Dockerfile
└── grammy/
    ├── bot.mjs             grammY Telegram bot + POST /send HTTP endpoint
    ├── package.json
    └── Dockerfile
```

## External Services (host ports, from openai/ stack)

| Service | Host Port | Role |
|---------|-----------|------|
| LiteLLM | 4000 | LLM proxy — all inference goes through here (`LITELLM_URL` env var) |
| Ollama GPU | 11436 | GPU inference backend + VRAM management (direct) + memory extraction |
| Ollama CPU | 11435 | nomic-embed-text embeddings for openmemory |
| Langfuse | 3200 | LLM observability — traces all requests via LiteLLM callbacks |
| Qdrant | 6333 | Vector store for memories |
| SearXNG | 11437 | Web search (used by `web_search` tool) |