sallm-agent

Minimal tool-calling agent (sallm) for local small LLMs (default: Gemma 4 4B via Ollama).

Focus: keep context predictable over long sessions — durable SQLite state, LanceDB retrieval, a skill stack, and a visible ContextReceipt that explains token spend.

Not an invisible RAG black box: raw messages stay canonical; vectors are a rebuildable index; derived facts must cite source message ids.

Setup

ollama pull gemma4:e4b-it-qat
ollama pull qwen3-embedding:0.6b

uv sync --extra dev

Core deps include peewee (SQLite ORM) and lancedb (vector index). There is no DSPy/Pydantic dependency.

Examples

• examples/imap_inbox/ — durable IMAP inbox Q&A (CLI tools + long-session recall). See the docstring in agent.py.

Turn pipeline

user → persist raw message
  → goal/skill control (small JSON call)
  → vector retrieve (Qwen embed + LanceDB)
  → budgeted prompt + ReAct ```run tools
  → persist answer
  → extract grounded facts + index chunks

Library

from sallm import Agent, RetrievalConfig, Skill, SkillRegistry
from sallm.tools import builtin_tools

agent = Agent(
    tools=builtin_tools(("calc", "echo")),
    state_path="/tmp/sallm/state.db",
    vector_path="/tmp/sallm/vectors",
    session_id="demo",
    retrieval=RetrievalConfig(
        memory_gate=True,
        search_mode="dense",  # or "hybrid"
        use_instruct=True,
        use_rewrite=False,
        use_hyde=False,
    ),
)
result = agent.ask("Remember the code is PURPLE-42.")
print(result["answer"])
print(result["receipt"])  # ContextReceipt as dict
print(result["goal"], result["stack"])

Resume by reusing state_path + session_id.

VectorStore contract

Implement upsert / search / delete_session / close (see sallm.memory.types.VectorStore). Default: LanceVectorStore. A future pgvector adapter can satisfy the same dataclasses (VectorRecord, VectorQuery, VectorHit) without changing the agent.

SQLite stores chunk text + indexed flags; LanceDB is rebuilt from those rows after a crash.

Skills

Default skill is converse. Register more with SkillRegistry (name, description, prompt fragment, optional tool subset).

Compiled profiles

Neutral JSON under sallm/profiles/ (instructions + demos + budgets). Offline:

uv run sallm optimize --dataset data/cases.jsonl --task controller --out /tmp/profile.json

sallm chat never optimizes at startup; it only loads a profile.

CLI

# Durable long session (recommended)
uv run sallm chat \
  --state-path .sallm/state.db \
  --vector-path .sallm/vectors \
  --session long1 \
  --retrieval-query instruct \
  --search dense \
  --memory-gate \
  --extract waterfall \
  --tools echo,calc

uv run sallm chat --show-prompt
uv run sallm chat --script tests/fixtures/sample_conversation.txt

Slash commands: /help, /clear, /history, /prompt, /state, /stack, /memory, /context, /quit.

Tool contract

Rule	Detail
Identity	Tool name = first argv token
Help	Every tool supports --help
Args	CLI flags only (no JSON blobs)
Intermediate	stdout may start with [intermediate]

```run
calc --expression "2**10"
```

Shipped tools: echo, calc, dig.

Legacy context optimizers

Still available without durable state: --context max-messages|summarize. Prefer --state-path + retrieval for hour-scale sessions.

Tests

uv run pytest tests/ -v
# E2E needs Ollama + gemma4:e4b-it-qat (+ qwen3-embedding:0.6b for stack memory)

Limits

Retrieval improves grounding; it does not guarantee the model never invents facts. Source-tagged memory and ContextReceipt make misses inspectable.

How it works

Walkthrough with an example, stage-by-stage flow, token-budget simulation, and hypothesis checks: docs/how-the-agent-works.md.

When a script turn (briefing / transcript) is larger than the history budget: docs/oversized-briefings.md.

Offline prompt/parameter tuning: docs/optimize-prompts.md.

Skills (selection, stack, tools): docs/skills.md.