Keep agent retrieval current as documents change with Serviette
Configure a watched document corpus, serve source-backed retrieval to agents, and verify that additions, edits, and deletions reach the index without manual re-ingestion.
npx skills add agentskillexchange/skills --skill keep-agent-retrieval-current-as-documents-change-with-serviette
Use Serviette when an agent needs current evidence from a changing document folder rather than a one-time upload. The operator selects a bounded corpus, configures its source, embedder and vector database in YAML, starts the indexer and retrieval server, then has an agent request relevant passages through the versioned REST API. Pathway watches source additions, edits and deletions and updates the vector store.
What this skill actually does
Install with Python 3.10 or later in a virtual environment using python -m pip install 'serviette[local]', with pip 23.1 or newer. The local extra installs a substantial PyTorch dependency stack; upstream documents a CPU-wheel index option. Obtain the required free Pathway license key separately and set PATHWAY_LICENSE_KEY. Run serviette wizard to generate config.yaml, select the document directory, local embeddings and the default DuckDB backend, then start serviette up --config config.yaml. Production deployments can run the documented indexer and server commands separately.
Query POST http://localhost:8989/api/v1/retrieve with JSON containing query and k. Return the retrieved text, source metadata and score for the agent to use as evidence. Verify freshness by asking a fixed question, changing a known fact in a test document, querying again, and checking that the source-backed result changes. Also test a document deletion so removed evidence is no longer offered. If evidence is absent or stale, report that limitation instead of inventing an answer. The optional /api/v1/rag endpoint returns generated answers and sources when its LLM configuration is supplied.
The scope is configuring and checking a live document-to-retrieval pipeline for an agent. It excludes building a generic chatbot product, inventing an MCP interface, model training, tenant management and unrestricted document collection. Multi-Framework is appropriate because any runtime able to make HTTP requests can invoke the same retrieval API and supply its own reasoning model. Local embeddings keep that stage local; choosing external embedding or answer providers sends the corresponding text to those providers. The required license signup and optional model credentials must be arranged by the operator. Upstream documentation and metadata were inspected; no installation, server deployment or security review was performed. Configuration and API documentation.