serviette

Looking for llm-app templates? All templates, cookbooks and examples are still available in full in the Pathway repository under examples/llm-app.

A no-code, always up-to-date RAG server for any vector database — built on the Pathway Live Data Framework.

Point serviette at a folder of documents, pick a vector database and an embedder in one YAML file, and run two commands: you get a chat and a REST API that answer from those documents and cite their sources. From then on, any change to the documents — an edit, a new file, a deletion — is reflected in the answers within seconds.

serviette runs on Pathway, which needs a free license key — one click at https://pathway.com/framework/get-license (GitHub or LinkedIn sign-in). Every command below expects it in PATHWAY_LICENSE_KEY.

Try it in four commands:

pip install serviette
export PATHWAY_LICENSE_KEY=...        # free: https://pathway.com/framework/get-license
export OPENAI_API_KEY=sk-...
serviette demo --embedder openai      # → chat at http://localhost:8989

Use pip 23.1 or newer (python -m pip install -U pip first): older pip lacks the resolver improvements this dependency tree needs and can take hours.

This indexes a small bundled corpus (a fictional coffee-machine company) into ./serviette-demo/docs and opens a chat over it. Edit or drop files into that folder while it runs — the answers follow within seconds.

One command to a live RAG stack — then edit a document and watch the answer change.

Who is it for

What you get. A question-answering assistant over your organisation's documents, with sources cited and answers that follow the documents: an edit, a new file or a deletion is reflected within seconds, with no re-upload or re-indexing step for anyone to run or forget.

What it costs to run. Two processes (an indexer and a server) and one YAML file. Embeddings can be computed locally for free; the Pathway license is free; the only metered component is the optional model API that writes the answer text, billed per question rather than per document.

Where your data goes. Nowhere by default. With the default local embedder, documents, their embeddings and the index stay on the machine or in the database you choose; the only thing that leaves is the handful of passages retrieved for a question, sent to the model API that writes the answer — and only if you configure one. (Choosing an API embedder instead sends the document text through that API as well.)

How mature it is. serviette is a thin layer over the Pathway Live Data Framework — the open-source engine that does the watching, parsing, embedding and writing: in development since 2022, 60 000+ GitHub stars, in production use at organisations such as DB Schenker, Intel, NATO, La Poste, Transdev and CMA CGM. serviette's benchmarks cover indexing up to 50 GB of text (26 million pages, 17 million files), with indexer memory measured throughout the run, and retrieval accuracy on a public benchmark (FRAMES). Eight vector databases are under integration tests against real instances. Scripts, recorded results and measurement limitations are available in this repository.

Where it stands today. It answers one question at a time, from the current state of the documents. Conversation memory, user accounts and a hosted offering are not built in yet — and each is a short step from what is here. If your use case needs one of them, or something else, contact us for a tailor-made solution: open an issue, write on Discord, or email contact@pathway.com.

The rest of this README is for the person setting it up.

To run it over your own documents, generate a config and start the stack:

pip install "serviette[local]" --extra-index-url https://download.pytorch.org/whl/cpu
    # [local] = free local embeddings, the wizard's default. It brings PyTorch; the
    # --extra-index-url picks its CPU build (~0.8 GB) instead of the default one
    # with CUDA libraries (~5 GB) — drop the flag only if you will embed on an
    # NVIDIA GPU. Plain `pip install serviette` if you will embed with OpenAI instead.
export PATHWAY_LICENSE_KEY=...   # the free Pathway license (see above); the wizard picks it up from here
export OPENAI_API_KEY=sk-...     # powers generated answers; omit to run keyless (answers quote the retrieved snippets)

serviette wizard                           # interactive config wizard
serviette up --config config.yaml          # indexer + server together → http://localhost:8989

(Production deployments run serviette indexer and serviette server separately — that is what up supervises.)

The server hosts both the web chat UI (on /) and the versioned REST API (under /api/v1) on one port:

curl -X POST http://localhost:8989/api/v1/retrieve \
  -H 'Content-Type: application/json' \
  -d '{"query": "how does persistence work?", "k": 5}'

Highlights

• No code. Configure everything in one YAML file (or generate it with serviette wizard).
• Any vector DB — 8 backends. DuckDB (embedded, zero setup — the default), pgvector, Qdrant, Milvus, ChromaDB, Weaviate, Pinecone and MongoDB Atlas Vector Search. Every backend is written through Pathway's native connectors, so file edits and deletions become real upserts/deletes in the store.
• Zero-setup default. DuckDB is the default backend: an embedded database in a single local file, no external service to install or run, with built-in vector search.
• Live & incremental. Built on Pathway: additions, edits and deletions are reflected in the vector DB in real time — whatever you change is answerable seconds later. Documents flow through the pipeline instead of accumulating in it, so a large corpus stays small in memory.
• Multiple sources. Local filesystem, Google Drive, S3/MinIO, SharePoint — plus anything the PyFilesystem library opens: FTP, SFTP, WebDAV, even ZIP archives. All watched live, mixed freely in one config.
• Multimodal out of the box. Text, Office documents, PDFs (with tables and layout), scanned images — and, with the corresponding API keys, audio recordings and even video. Every format is on by default and routed to the best parser that needs no API key; drop a file in the folder and it is answerable like any document. See Multimodality.
• Reuses Pathway's LLM xpack. Parsers, splitters and embedders are used as-is — serviette implements none of its own. Five embedder families (OpenAI, LiteLLM, SentenceTransformers, Gemini, Bedrock) work identically on the indexer and the server side — including a fully local, credential-free stack with local embeddings + DuckDB.
• Decoupled & scalable. Indexer and API server are independent processes sharing only the vector DB. The server is stateless and scales horizontally; the indexer shards across worker processes with one config line. Every part scales on its own.
• Web chat UI, same port. serviette server serves a clean ChatGPT/Claude-style chat page on / next to the versioned API (/api/v1/...) — same origin, no CORS, nothing extra to run. For split deployments (UI on a different host) there is a standalone serviette frontend proxy tier.
• Free Pathway license. One click at https://pathway.com/framework/get-license.

Architecture

The two halves are fully decoupled. The indexer (write path) and the server (read path) are separate processes — different executables that never talk to each other. Their only contract is the vector database itself:

• Independent scaling. The server is stateless and read-only — run any number of instances behind a load balancer; each also serves the chat UI at zero cost. The indexer scales separately: Pathway shards it across worker processes (indexer.workers: 8 is how the benchmarks below run), so every part of the stack scales independently. Bulk re-indexing never slows down query serving, and query spikes never stall indexing.
• Failure isolation. If the indexer is down, serving continues over the last-synced data; if the server is down, indexing keeps the database fresh. Either side can be restarted or upgraded independently (the indexer resumes from its persistence without re-embedding).
• The database stays yours. Vectors live in your store in a plain, documented schema — other consumers (BI, other apps, a different retrieval stack) can read the same collection; serviette doesn't hold it hostage. And since the default store is an embedded DuckDB file, trying this out costs nothing to set up.
• Optional third tier. For split deployments (UI on a different host than the API) a standalone serviette frontend serves the same chat page and proxies to the API server-side.

The one deliberate exception: the embedded DuckDB backend trades this distribution for zero setup — one local file, single-writer, ideal for laptops and demos (see docs for its concurrency note).

Benchmarks

Two self-contained benchmarks, one per axis: what indexing costs in time and memory, and how accurate the retrieval is.

Indexing resources

Self-contained benchmark (docker-compose: Qdrant + indexer + server, fully local embeddings, zero API cost) over a Wikipedia corpus of plain text — every byte below is extracted text (a PDF collection with the same text content would weigh several times more) — see benchmarks/realtime-data-indexing:

corpus target	≈ pages	files	chunks (approx.)	indexing time	peak memory (PSS, GiB)	in Qdrant (GiB)
100 MB	52 431	12 969	53 651	44 s	6.86	0.61
1 GB	524 288	240 516	604 299	4.9 min	6.82	1.85
3 GB	1 572 864	841 890	1 888 846	15.4 min	7.18	4.89
10 GB	5 242 884	3 423 359	6 761 214	1.0 h	8.16	16.07
30 GB	15 728 646	9 202 620	20 933 962	3.1 h	10.40	48.57
50 GB	26 214 382	17 083 603	37 472 786	6.0 h	14.65	84.45

The table and both figures come from the same six runs. Raw summaries, memory samples, per-process snapshots and resolved configurations are in results. One final run per size is published; run-to-run variability was not measured. Memory and disk sizes use GiB; input labels are nominal targets (1 GB = 1,000 MiB of extracted text). A page is approximately 2,000 bytes. Chunks are Qdrant's approximate count at completion; the per-size curves retain the maximum observed count used by the completion detector.

Documents flow through the pipeline rather than accumulating in it. The file-watch index retains metadata per watched file to detect edits and deletions, so its cost depends on file count. The right panel estimates this cost from the connector worker's extra private memory: the fitted slope is ~361 bytes per file. This proxy also includes unequal worker buffers; it does not isolate individual allocations.

The embedding runtime, engine and working buffers account for the worker baseline. The curves show how total memory evolves during indexing, including any steps as file-watch metadata and buffers grow. Parsed-text cache, persistence snapshots and stored embeddings live on disk. Across these runs a 500× larger input target costs 2.1× the peak PSS. Both file count and text volume change between sizes; this comparison does not establish constant memory usage.

All curves and table values come from the same runs with allocator background purging enabled. File-watch memory is estimated from rank-0 private-memory excess over the median of the other workers.

Attribution at the sampled 10 GB peak. Local-embedder memory is estimated from a separate eight-process initialization diagnostic and library mappings; libraries already needed without the embedder remain in the residual. The 3.27 GiB estimate without the local embedder is calculated by subtraction, not measured in an external-embedder indexing run. It excludes the embedding service and does not measure client overhead or changes in working buffers. Shared pages are accounted for proportionally through PSS.

Setup: 96-logical-CPU host (see the saved CPU model), streaming mode, 8 worker processes with 12 OpenMP threads each, local static-retrieval-mrl-en-v1 embeddings, Matryoshka-truncated to 256 dims, 512-token chunks, Qdrant, and jemalloc background_thread:true. No embedding API calls. Installed versions: serviette 0.1.7, Pathway 0.33.0, PyTorch 2.14.1, sentence-transformers 6.1.0. Peak memory is the maximum sampled PSS summed over the indexer container. Qdrant and API-server memory, unmapped filesystem cache and kernel memory are outside that measurement; it is not a RAM requirement for the whole stack. Brief peaks between samples may be missed.

Indexing time is an estimate ending at the last increase in the running maximum of Qdrant's approximate chunk count. That moment can precede the final write; output-completeness checks do not revise the timestamp or measure when the backend's search index is ready. The recorded environment, cache conditions and limits on reproducing it are described in the measurement notes.

Retrieval accuracy (FRAMES)

End-to-end evaluation on FRAMES (Google, 2024): 824 multi-hop questions whose answers must be assembled from 2–15 English Wikipedia articles — see benchmarks/frames, full technical report in REPORT.md:

measurement	result
gold-article recall — the paper's metric, on the paper's corpus	0.50 vs 0.15 published for the paper's BM25 baseline (0.21 for our reproduction of it)
paired gain from adding serviette to gpt-5, paper's protocol	+5.2 pp over the same model without retrieval (McNemar z = 4.1, 824 questions)
adaptive retrieval in the grounded (context-only) regime	41.3% → 52.8% (z = 7.5) — the largest single effect measured
absolute accuracy (permissive, gpt-5)	73.7% — above every number in the paper, including its 5-step agent (66.0%) and oracle (72.9%)

The setup reproduces the paper wherever technically possible: the identical Wikipedia dump (TFDS wikipedia/20230601.en, 5.22M articles → 12.07M chunks) indexed in full by serviette with the free local e5-small-v2 embedder — so the recall row costs nothing in API fees — plus the paper's own retrieval metric and its verbatim autorater prompt. Comparisons are paired and internal (identical corpus, generator, judge; only the retrieval layer varies — the opt-in strategies described under Retrieval quality below), because the absolute score is dominated by the 2026 generator — its no-retrieval baseline alone reaches 68.5% — which is why the headline is the paired delta, not 73.7%. The grounded rows measure serviette as it ships for private corpora: answers strictly from retrieved documents, a deliberately stricter regime than the paper's. Methodology, statistics, limitations and raw per-question outputs: REPORT.md.

Multimodality

Every file type is enabled by default. serviette routes each file to the best parser that works without an API key, and turns on key-requiring modalities automatically when their key is present:

format	parsed by default with	notes
text / Markdown	as-is
PDF	pypdf — built in; serviette[docling] upgrades to Docling (layout-aware, tables)	local, free
Office (DOCX, PPTX, XLSX, HTML, EML…)	Unstructured — built in; serviette[docling] widens coverage (EPUB, legacy formats)	local, free
scanned images (PNG, JPG, TIFF…)	PaddleOCR	local, free; install serviette[ocr]
audio (MP3, WAV…)	Whisper	when OPENAI_API_KEY is set
video (MP4, WebM, MOV…)	TwelveLabs Pegasus — a searchable text description of the video	when TWELVELABS_API_KEY is set

A modality whose only parser needs an absent key is skipped with a clear warning — never a crash. Everything stays live: drop a recording of yesterday's meeting into the watched folder and ask about it minutes later; expensive parses (video) are cached on disk, so restarts cost nothing.

The routing is configurable per file pattern (parser: section — pick a vision model for images instead of OCR, set a custom video prompt); see docs. Embeddings work the same for every modality: parsed content is text, so any of the embedder families — including the local credential-free default — covers a multimodal corpus.

Retrieval quality

Retrieval and answering are pure-vector by default; five opt-in strategies improve accuracy, each a few lines of config. They compose freely — decompose widens what is retrieved, hybrid and reranking reorder which chunks win, MMR diversifies them, and adaptive RAG retries with more context when the answer is not found.

• Hybrid (BM25 + vector). Fuses vector similarity with an in-process BM25 keyword index (reciprocal-rank fusion). The two signals fail differently — embeddings match paraphrases, BM25's IDF makes rare exact tokens (names, dates, IDs) dominate — so their fusion beats either alone on entity-heavy corpora. Enable with hybrid: true on the vector-db section.
• Reranker. A second stage rescoring a shortlist of candidates: a local, keyless cross-encoder (reranker: {type: cross_encoder}) or pointwise LLM scoring (type: llm). Best for precise, single-hop questions.
• Adaptive RAG. Answers from k chunks first, then grows the context and re-asks while the LLM reports the answer is not present (rag.adaptive). Your llm.system_prompt stays in force; the mode only appends the "reply with this marker when the context is insufficient" instruction it needs to decide when to fetch more.
• Query decomposition. One LLM call splits a multi-hop question into sub-queries, retrieves for each, and fuses — so chunks of different hops stop competing for the same top-k slots (rag.decompose).
• MMR. Diversifies the result set, trading relevance against redundancy (rag.mmr).
• Document questions. "How many documents do you have?", "What does journal.pdf say?", "What changed between law_2023.txt and law_2024.txt?" are questions about documents, not about a passage, and chunk retrieval cannot answer them. rag.documents (on by default) answers them from the document catalog and from whole files: a question naming a file goes straight to that file; any other question is searched as usual, and only if that finds nothing does one short LLM call decide between a wider search and the catalog or the named documents. Two files that share most of their text are compared as a line-level diff computed by the server, so the model reads only what changed — however the files were chunked. Every step has a fixed context budget and the request as a whole a ceiling (max_request_chars); when a file cannot be identified the answer says which names were close. Not available on Pinecone, which cannot enumerate documents.

The reranker and the multi-step strategies (adaptive, decompose) are backend-independent — they work on every vector DB. Hybrid and MMR read from the store, so support depends on the backend:

Backend	Vector	Reranker · Adaptive · Decompose	MMR	Hybrid (BM25)
DuckDB	✅	✅	✅	✅ in-process
Qdrant	✅	✅	✅	✅ in-process
pgvector	✅	✅	✅	✅ in-process
Milvus	✅	✅	✅	✅ in-process
Weaviate	✅	✅	✅	✅ in-process
ChromaDB	✅	✅	✅	✅ in-process
MongoDB	✅	✅	✅	✅ in-process
Pinecone	✅	✅	✅	❌ — no scan-all API; native sparse-index hybrid is planned

The in-process index follows the store: it is rebuilt as soon as the backend reports a change (the row count; on DuckDB and pgvector also the newest seen_at, which catches in-place edits) and refreshed in the background every hybrid_refresh_seconds (default 30) as a bound on staleness for whatever that cheap signal misses. In-process BM25 targets corpora up to a few million chunks; above hybrid_max_chunks the keyword leg is skipped with a warning and retrieval stays pure-vector. Pinecone cannot enumerate its vectors, so it has no in-process hybrid — hybrid: true there fails fast at startup with that explanation; native hybrid (a second sparse index) is planned. Native server-side BM25 for the client-server backends (replacing the in-process leg at larger scale) is planned as well.

Observability

Three layers, all on by default or one config line away:

In the chat UI. The header shows "indexed N s ago" — the age of the most recent write into the vector store. When you edit a source document, you can watch the counter reset as the change lands.

GET /api/v1/stats on the API server: the backend in use, the number of indexed chunks, and index freshness — a JSON one-liner for dashboards and health checks, served without touching the indexer (it reads the vector store, like every other query).

Engine metrics (Prometheus). The Pathway engine ships its own observability server; serviette exposes it with one config line:

indexer:
  monitoring_http_port: 20000

Every worker process then serves GET /metrics on 127.0.0.1:(20000 + worker index) — input/output latency gauges (i.e. the indexing lag behind the sources) and per-operator row counters, straight from the engine's dataflow. Point a Prometheus scrape at the worker ports and you get per-stage throughput and freshness graphs with no extra code.

Logs from both processes go to stdout/stderr in plain text; serviette up interleaves them with per-process prefixes.

Security

The server listens on localhost only by default and ships no built-in authentication — exposing it is an explicit decision: set server.host: 0.0.0.0 and put an authenticating reverse proxy in front (a five-line Caddy example lives in docs).

Requirements

• Python ≥ 3.10 (the minimum supported by Pathway).
• A free Pathway license key in PATHWAY_LICENSE_KEY — https://pathway.com/framework/get-license.

Documentation

Full installation, the config wizard, configuration reference, persistence, architecture and scaling notes live in docs/README.md.

Development

To report a problem or propose a change, see CONTRIBUTING.md for the contribution policy and PR requirements.

serviette runs on the released Pathway from PyPI (≥ 0.32.1 — the first release with the vector-database connectors). From-scratch setup on a fresh machine:

# 0. Prerequisites: Python >= 3.10 (3.12 recommended) and git.

# 1. serviette in its own virtualenv
git clone https://github.com/pathwaycom/serviette.git && cd serviette
python3.12 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,local]"

# 2. A (free) Pathway license: https://pathway.com/framework/get-license
export PATHWAY_LICENSE_KEY=...

# 3. Sanity check: zero-to-chat on the bundled corpus
serviette demo                        # -> http://localhost:8989

Notes:

• serviette demo materializes everything in a visible working directory: ./serviette-demo/docs holds the corpus — a toy example, a handful of documents about a fictional company (Lumina Coffee Systems). Drop your own files there (PDF, DOCX, scans, …) while it runs and they are answerable within seconds.
• The demo indexes with one of two embedders — pick your trade-off:
  • default, free & local — needs serviette[local], which pulls the PyTorch stack. Install it with --extra-index-url https://download.pytorch.org/whl/cpu to get the CPU build (~0.8 GB); without the flag pip takes PyTorch's default Linux wheel, which ships CUDA libraries whether or not you have an NVIDIA GPU (~5 GB, and the install is the slow part of the first run). Free at any corpus size afterwards.
  • --embedder openai — nothing to install, starts immediately, but every indexed token is billed to your OPENAI_API_KEY. Fine for the toy corpus and small folders; for a large collection, sit out the serviette[local] install and use the free embedder instead.
• Independently of the embedder, export OPENAI_API_KEY if you want real generated answers in /rag — without it the demo answers by quoting the retrieved snippets.

Running the test suites:

pytest -m "not slow"            # fast unit tests (no Pathway, no services)
pytest -m "slow and not integration"   # end-to-end indexer tests (spin up Pathway)
pytest -m integration          # real-database tests (see below)
pytest                         # everything

Per-backend clients install as extras — pick what you use:

pip install "serviette[qdrant]"      # also: pgvector, milvus, chroma, weaviate,
                                  #       pinecone, mongodb, local, gemini, all

Integration tests (real databases)

Every claimed backend has an integration test running the same scenario end-to-end against a real instance: index two documents with the real indexer, retrieve through the production accessor (exact-text query must rank first with cosine ~1.0), delete a file, re-index, and verify its vectors are gone (snapshot semantics). The shared driver lives in tests/integration_common.py.

Backend	Test	Real instance
DuckDB	test_integration_duckdb.py	embedded — runs everywhere
pgvector	test_integration_pgvector.py	pgvector/pgvector Docker container
Milvus	test_integration_milvus.py	embedded Milvus Lite engine; set SERVIETTE_TEST_MILVUS_URI=http://host:19530 to run the same tests against a Milvus server (Lite has no growing/sealed segments, so it cannot catch stats-counting bugs)
Qdrant	test_integration_qdrant.py	qdrant/qdrant Docker container
ChromaDB	test_integration_chroma.py	chromadb/chroma Docker container
Weaviate	test_integration_weaviate.py	semitechnologies/weaviate Docker container
Pinecone	test_integration_pinecone.py	official pinecone-local emulator (Docker)
MongoDB	test_integration_mongodb.py	mongodb-atlas-local (mongod + mongot, real $vectorSearch)

Containers are throwaway (tests/dockerutil.py, Docker CLI via subprocess, no extra dependency) and host ports are allocated dynamically — tests never assume a fixed localhost port is free or that a service is already running. Each test skips automatically when Docker or its client library is missing.

License

See LICENSE.