Dynamic Context Discovery

Artifact-based storage for large tool outputs

from praisonaiagents import Agent

agent = Agent(
    name="researcher",
    instructions="Use artifact tools when tool output is large.",
)
agent.start("Fetch the full API schema and summarise the auth section.")

Large tool outputs are queued to artifacts automatically — the agent sees compact references instead of flooding context, and explores data on demand with artifact tools.

The user asks for large data; middleware stores bulky tool output as artifacts and keeps compact references in context.

Quick Start

1. Enable dynamic context on an agent

   from praisonaiagents import Agent
   from praisonai.context import setup_dynamic_context
   
   ctx = setup_dynamic_context()
   
   agent = Agent(
       name="Analyst",
       instructions="You are a data analyst.",
       tools=ctx.get_tools(),
       hooks=[ctx.get_middleware()],
   )
   
   agent.start("Fetch and analyse the large dataset")
   

   Outputs over 32 KB are saved as artifacts; the agent receives references and uses artifact_tail, artifact_grep, etc.

2. Tune the inline threshold

   ctx = setup_dynamic_context(
       inline_max_kb=16,
       redact_secrets=True,
       history_enabled=True,
   )
   

---

How It Works

Outputs above inline_max_kb are written to disk under base_dir. The agent sees a short reference and pulls only the lines it needs via artifact tools — context stays bounded even after heavy tool runs.

---

Available Tools

Tool	Description
artifact_tail	Last N lines
artifact_head	First N lines
artifact_grep	Search for pattern
artifact_chunk	Line range
artifact_list	List artifacts

Configuration

Parameter	Default	Description
inline_max_kb	32	Queue outputs larger than this
redact_secrets	True	Redact API keys and passwords
history_enabled	True	Persist conversation history
base_dir	~/.praisonai/runs	Artifact storage directory

Environment variables: PRAISONAI_ARTIFACT_DIR, PRAISONAI_ARTIFACT_INLINE_MAX_KB, PRAISONAI_ARTIFACT_REDACT.

Best Practices

Lower inline_max_kb for memory-heavy agents

16 KB keeps more headroom in context for long multi-tool sessions.

Keep redact_secrets enabled in production

Automatic redaction prevents credentials in tool output from reaching the LLM or disk logs.

Use artifact_grep before loading full files

Search artifacts first — avoid pulling entire multi-MB outputs into a single prompt.

Related

Context Management

Full context system

Security & Redaction

Secret redaction details