py-basic-agent

A tiny but real agentic system you can run locally. It uses:

• an LLM (via Ollama) for reasoning and tool selection
• pgvector for long‑term memory (document embeddings in Postgres)
• a Rich + prompt_toolkit REPL for a pleasant CLI
• optional MCP (Model Context Protocol) tools (HTTP façade) for file I/O
• a minimal ETL pipeline (CSV/JSON → transform → save), plus simple web search/fetch tools

RAG support: The REPL now includes a small Retrieval‑Augmented Generation flow. See RAG (Retrieval‑Augmented Generation) below.

---

Contents

• Quick start
• Requirements
• Setup (Linux / macOS / Windows)
• Ollama hosting
• Environment variables
• Reset after embedding update
• How it works (core concepts)
• Using the REPL
• ETL mini‑DSL
• MCP (Model Context Protocol) tools
• RAG (Retrieval‑Augmented Generation)
• Troubleshooting
• Project layout

---

Quick start

Before you start: make sure Ollama is running and reachable. See Ollama hosting for OS‑specific setup (macOS runs Ollama outside Docker; Linux/Windows may use the optional ollama service).

# 1) Start dependencies (DB + optional MCP file server)
docker compose up -d pgvector mcpfs

# 2) Run the agent REPL
docker compose run --rm app

In the REPL:

/research Who founded NVIDIA and when?
/etl -p ./data/sales_orders.csv -t "reorder:date,region,product,units,unit_price; rename:unit_price->price; limit:3"
/mcp add-http -n fs -u http://host.docker.internal:8765
/mcp tools
/mcp call fs list_files '{"path":"./data"}'

Type /help for a full list of commands, or exit() to quit.

RAG quick taste:

/rag ingest -p ./knowledge
/rag show -q "quartz-8127" -k 3
/rag ask What is the RAG demo code?

---

Requirements

You don’t need Python locally. Everything runs in containers.

• Docker Desktop or Engine (compose v2)

• Ollama running on your host (for the LLM + embeddings)

  • pull models you’ll use:

    ollama pull llama3.1:8b
    ollama pull all-minilm
    

  • make sure Ollama is reachable from Docker at http://host.docker.internal:11434

• ~2–4 GB free RAM for the small models

---

Setup (Linux / macOS / Windows)

1) Clone and prepare .env

git clone <your-repo-url>
cd py-basic-agent
cp .env.example .env

Update values as needed (see Environment variables).

2) Start Postgres (pgvector) and MCP file server

docker compose up -d pgvector mcpfs

The pgvector container auto‑creates the database, role, and tables via the mounted db/schema.sql. No manual DB bootstrapping needed.

3) Run the REPL

docker compose run --rm app

(Optional) Prepare a knowledge folder for RAG

Create knowledge/ with demo docs; you can use the included intro.md and policies.md examples.

mkdir -p knowledge
# these files may already exist in the repo
# echo "# RAG Demo — py-basic-agent" > knowledge/intro.md
# echo "# PII Handling & Redaction Policy (Demo)" > knowledge/policies.md

Mount it in compose (see RAG section) or use a repo‑relative path with /rag ingest -p ./knowledge.

---

Ollama hosting

• macOS (recommended): run Ollama outside Docker (host app). This is required on macOS so the models can use Metal and to avoid Docker networking issues. Keep OLLAMA_HOST=http://host.docker.internal:11434.
• Linux/Windows (optional in‑container): you may run Ollama as a container with Compose. When running the in‑container service, set OLLAMA_HOST=http://ollama:11434.

⚠️ Experimental: the dockerized Ollama path depends on GPU drivers/permissions. If you hit timeouts or slow responses, prefer running Ollama on the host and use host.docker.internal:11434.

---

Environment variables

All live in .env. Defaults are safe for local use.

Variable	What it does	Example
OLLAMA_HOST	Base URL for Ollama. The app calls /api/chat and /api/embeddings.	http://host.docker.internal:11434
AGENT_LLM_MODEL	Chat model used by the agent.	llama3.1:8b
AGENT_TEMPERATURE	LLM sampling temperature (float).	0.2
AGENT_MAX_STEPS	Agent loop max tool steps before giving up.	8
AGENT_DB_URL	Postgres URL used by PgVectorMemory. Provided by compose.	postgresql://agent:agentpass@pgvector:5432/agentdb
AGENT_EMBED_MODEL	Embedding model name in Ollama.	all-minilm
AGENT_EMBED_DIM	Embedding vector dimension. Must match schema.sql.	384
SERPER_API_KEY	API key for the Serper search tool (if enabled).	sk-…
AGENT_VERBOSE	Controls extra logging in some contexts (true/false).	true
KB_PATH	Default folder for RAG ingestion inside the container. If you don’t mount it, use /rag ingest -p ./knowledge.	/knowledge

Changing models

• Use a different chat model: set AGENT_LLM_MODEL (e.g., llama3.2:3b).
• Use a different embedding model: set AGENT_EMBED_MODEL and update AGENT_EMBED_DIM, then update the DB schema (vector size in schema.sql) and recreate DB or migrate:

ALTER TABLE docs ALTER COLUMN embedding TYPE vector(<new_dim>);

---

Reset after embedding update

If you change AGENT_EMBED_MODEL or AGENT_EMBED_DIM, ensure the database vector dimension matches the embedding dimension. Easiest path is to recreate the DB volume (auto‑initialized from db/schema.sql).

1. Edit .env to the new model & dim (defaults use all-minilm).
2. Edit db/schema.sql to match (e.g., VECTOR(384)).
3. Drop and rebuild Postgres (recreates volume and schema):

docker compose down -v
docker compose up -d pgvector

1. (Optional) verify the table shape:

docker compose exec -T pgvector psql -X -U agent -d agentdb -c "\d+ docs"

1. Restart the REPL:

docker compose run --rm app

Prefer a migration instead of dropping the volume? Use the ALTER TABLE … TYPE vector(<dim>) command above.

---

How it works (core concepts)

1) The Agent loop (agent/core.py)

• Builds a message list: a system prompt + user task + any observations from tools.
• Calls the LLM with the messages.
• The LLM responds with a JSON tool call or a final answer, e.g.: { "tool": "search", "input": {"query": "NVIDIA founders"} }.
• The agent executes the tool, captures an observation (and often a short LLM summary), appends that back to the conversation, and repeats until a final answer or the step limit is reached.

2) The LLM client (agent/llm.py)

• Thin async wrapper around Ollama APIs:

  • /api/chat for conversation
  • /api/embeddings for vectorization

• Normalizes messages and enforces length limits.

• Includes a system prompt that teaches the LLM to return JSON tool calls or a final JSON with a readable summary.

3) Tools (agent/tools.py)

• Search: Serper (or stub). Returns top results.
• Fetch: Gets a URL, returns {title, url, text}.
• ETL: load_csv/json → transform (select/rename/limit) → save.
• Memory: router that calls the configured memory implementation.

Each tool is pure async and returns structured data.

4) Memory (agent/memory/)

Two implementations (same interface):

• SimpleMemory (in‑process, substring search) — great for unit tests.

• PgVectorMemory (Postgres + pgvector, cosine similarity):

  • aupsert(docs) store content + embeddings
  • aquery(query, k) vector search
  • aadd(text, source, uri, meta) quick note
  • adump(n) latest notes dump (for system prompt context)

Embeddings come from Ollama (AGENT_EMBED_MODEL), and the vector size must match DB schema.

5) REPL (agent/repl.py)

• Rich header + tool hints + keybindings

• History, completion for commands and file paths

• Commands:

  • /research <question>
  • /etl -p <path> -t "<transform>" [-l <out>]
  • /etl_from_source -p <url> -t "<transform>" [-l <out>]
  • /where <path>
  • /mcp … (see below)
  • /help, exit()
  • /rag … (see below)

---

Using the REPL

Start it:

docker compose run --rm app

Try research:

/research Compare NVIDIA vs AMD GPU market share over the last 2 years.

Try ETL:

/etl -p ./data/sales_orders.csv -t "reorder:date,region,product,units,unit_price; rename:unit_price->price; limit:3"

Check a path:

/where ./data/sales_orders.csv

---

ETL mini‑DSL

Chain operations with semicolons. Works for CSV columns and JSON keys.

• reorder:colA,colB,colC — Reorder columns; unspecified columns are appended in original order.
• rename:old1->new1,old2->new2 — Rename fields/columns. Quote names with spaces or numeric keys: rename:'1958'->y1958,'unit price'->price.
• limit:K — Truncate rows/objects to K.

Examples:

/etl -p ./data/sales_orders.csv -t "reorder:date,region,product,units,unit_price; rename:unit_price->price; limit:3" -l ./data/sales_orders.sample.parquet

---

MCP (Model Context Protocol) tools

MCP lets you add external tools to the agent at runtime (HTTP façade or stdio servers).

Examples:

/mcp add-http -n fs -u http://host.docker.internal:8765
/mcp tools
/mcp call fs list_files '{"path":"./data"}'

---

RAG (Retrieval‑Augmented Generation)

A small, readable RAG layer is built into the REPL so you can demonstrate how articles/notes become ground truth for answers.

Files & mounting

• Place .md and .txt files under knowledge/ in your repo.

• Either:

  • run: /rag ingest -p ./knowledge (no compose changes), or

  • mount the folder into the container and use the default KB_PATH:

    # compose.yml (app service excerpt)
    services:
      app:
        environment:
          KB_PATH: /knowledge
        volumes:
          - ./:/app
          - ./knowledge:/knowledge:ro
    

Commands

/rag ingest [-p PATH] [--glob "*.md,*.txt"]
/rag add -t "text" [-s source] [-u uri]
/rag show -q "query" [-k 6]
/rag ask  <question> [-k 6]

• ingest — indexes files into pgvector (chunks ≈ 800 words, 150 overlap).
• add — stores a one‑off snippet as a doc chunk.
• show — vector search only; pretty prints sources/URIs and highlights terms.
• ask — retrieves context and asks the LLM to answer only from that context (responds “I don’t know.” when not covered).

Demo knowledge

The repo includes knowledge/intro.md and knowledge/policies.md with respective test tokens:

• RAG‑DEMO-INTRO‑CODE: quartz-8127

• RAG‑PII‑POLICY‑CODE: heron-4512

PII behavior to demo:

/rag ask What is Jane Roe's SSN?
# → Per policy, I cannot share that information. [REDACTED]

/rag ask What is the PII policy code?
# → heron-4512

Smoke checks

/rag ingest
/rag show -q "quartz-8127" -k 3
/rag ask What is the RAG demo code?

Expect: ingest counts > 0, retrieval shows intro.md, answer includes quartz-8127.

Notes

• Embedding dim must match DB schema (VECTOR(dim)); change AGENT_EMBED_MODEL/AGENT_EMBED_DIM together and re‑ingest.

• Apostrophes are safe in /rag ask (parser avoids shlex pitfalls).

• If /rag ingest returns files=0, verify mounts:

  • /where /knowledge vs /where ./knowledge.

---

Troubleshooting

• Ollama not reachable: check OLLAMA_HOST; test with curl $OLLAMA_HOST/api/tags from inside a container.
• No embeddings / zero scores: ensure the embed model is pulled (all-minilm) and dims match schema.
• RAG ingest finds zero files: mount ./knowledge:/knowledge:ro or pass -p ./knowledge.
• Docker networking: on macOS/Windows use host.docker.internal; on Linux, see extra_hosts in compose.

---

Project layout

app/
  agent/
    core.py         # agent loop
    llm.py          # Ollama chat/embeddings client
    repl.py         # REPL (commands include research/etl/mcp/rag)
    tools.py        # tool registry (search/fetch/etl/memory)
    memory/
      __init__.py   # get_memory() factory
      pg_store.py   # PgVectorMemory impl
      simple.py     # SimpleMemory impl
knowledge/
  intro.md          # RAG demo document (token quartz-8127)
  policies.md       # optional PII policy demo (token heron-4512)
compose.yml
.env.example
README.md