A DataFrame layer that stops agents drowning in their own context

What it does

fenic acts as a pre-processing layer between raw data and your agent. You feed it sources like PDFs or chat logs, then use a DataFrame-style API to apply deterministic filters alongside semantic transforms—parsing, embedding, structured extraction, summarization—to build typed tables. These tables are exposed to agents as narrow MCP tools or Python functions, so the runtime receives small, precise rows instead of raw documents or bloated conversation histories.

The interesting bit

The project treats context engineering as a data pipeline rather than a prompt-hacking exercise. By wrapping LLM-powered operations in declarative DataFrame operations, it moves token-heavy work out of the agent’s context window, letting the agent spend its limited attention on reasoning instead of data janitoring.

Key highlights

• Mixes deterministic SQL-like ops (filter, join, unnest) with semantic primitives (semantic.extract, semantic.embed, semantic.parse_pdf) in one composable flow.
• Framework-agnostic output: serves shaped context via MCP servers or direct Python functions to LangGraph, PydanticAI, CrewAI, or anything else.
• Inference offloading: LLM calls for summarization, extraction, and embedding happen outside the agent’s prompt, reducing context bloat.
• Built-in memory patterns: curated facts, profile blocks with timelines, decaying temporal resolution, and cross-agent shared tables.
• Typed, bounded tools: outputs conform to Pydantic schemas, giving agents a narrow, predictable surface instead of open-ended document access.

Caveats

• The README examples import from internal-looking modules such as fenic.api.mcp._tool_generation_utils and fenic.core.mcp._server, suggesting the API may still be settling.
• It offloads inference from the agent’s window but does not eliminate LLM costs—you still pay for the tokens fenic consumes during extraction and embedding.
• How save_as_table persists data and what backends are supported is not explained in the README.

Verdict

Worth evaluating if your agents struggle with large documents or long conversational memory and you want a structured pre-processing layer. Skip it if you need a full agent runtime—fenic is explicitly glue, not a framework.