Eine lokale, GPU-beschleunigte Retrieval-Pipeline für Forschungspapiere – parst PDFs mit layoutbewusstem Chunking, zieht echte bibliografische Metadaten (DOI → CrossRef), embeddet mit hybriden dichten + spärlichen Vektoren und speichert alles in Qdrant für hybride Suche mit Reranking. Entwickelt, um direkt von Claude über einen MCP-Server abgefragt zu werden.
Papiere in einen Ordner zu legen macht sie für die Suche unsichtbar. Das verwandelt einen Stapel PDFs in etwas Abfragbares: saubere Abschnitts-/Absatz-bewusste Chunks (kein naiver Zeichen-Splitter), zitierfertige Metadaten, die automatisch von CrossRef geholt werden, und hybride dichte+spärliche Retrieval, sodass sowohl semantische als auch exakte Begriffssuchen funktionieren – alles als MCP-Tools exponiert, damit Claude Papiere direkt durchsuchen und aufnehmen kann.
Related MCP server: Personal Research Assistant MCP
Layoutbewusstes PDF-Parsing über Docling – chunked nach Absatz/Abschnitt statt mit einem naiven Zeichen-Splitter, behält Seitennummern und Abschnittsüberschriften pro Chunk.
Echte bibliografische Metadaten – extrahiert per Regex eine DOI von Seite 1, löst vollständige Zitationsdaten (Titel, Autoren, Zeitschrift, Jahr, Band/Ausgabe/Seiten) über die CrossRef-API auf, fällt auf die eigenen PDF-Metadaten zurück, wenn keine DOI gefunden wird.
Hybride Embeddings – dichte (BAAI/bge-m3) + spärliche (SPLADE, prithvida/Splade_PP_EN_v1) Vektoren pro Chunk, gemeinsam in Qdrant für hybride Retrieval gespeichert.
Hintergrund-Ingestion – ein Redis + RQ-Worker, damit große PDFs keine Anfrage blockieren; GPU-fähiger Docker-Container für Embedding.
MCP-Server, der Claude zwei Tools exponiert – search_papers (hybride Suche mit Reranking) und ingest_paper (ein PDF direkt in den Index werfen) – gedacht als Forschungs-Papier-Gedächtnis, das Claude direkt abfragen kann.
Dockerisiert – Qdrant + Redis + GPU-fähige API/Worker-Container über docker-compose.
Python · Docling (PDF-Parsing) · sentence-transformers (BGE-M3 dichte Embeddings) · fastembed (SPLADE spärliche Embeddings) · Qdrant (hybride Vektorsuche) · Redis + RQ (Hintergrund-Job-Warteschlange) · FastAPI (geplante API-Schicht) · MCP (Model Context Protocol-Server für Claude) · Docker Compose, CUDA 12.9 GPU-Container.
In Arbeit – noch nicht vollständig Ende-zu-Ende verdrahtet.
Funktioniert heute: ingest.py läuft eigenständig von der CLI – ein PDF parsen, seine Metadaten ziehen, embedden (dicht + spärlich) und in Qdrant upserten. Auf eine laufende Qdrant-Instanz zeigen und es funktioniert.
Fehlt: main.py, der FastAPI-Dienst, der /search, /ingest und /health exponieren soll – das Dockerfile kopiert ihn, der Docker-Compose-Healthcheck pingt ihn an, der RQ-Worker erwartet, dass etwas Jobs für ihn einreiht, und die beiden Claude-Tools von mcp_server.py rufen ihn beide über HTTP unter localhost:8000 auf. Ohne ihn hat der Worker nichts zu konsumieren, die MCP-Tools haben kein Backend, mit dem sie sprechen können, und docker-compose up wird nicht bauen (der Build-Schritt des API/Worker-Images kopiert eine Datei, die nicht existiert).
Die Bind-Mounts in der docker-compose.yml zeigen außerdem noch auf einen Windows-Pfad (S:\RAG-data\...) aus einem früheren Setup – passe diese an, wo auch immer Qdrant/Redis/Modell-Cache-Daten leben sollen, bevor du es ausführst.
uv sync
# Qdrant needs to be running somewhere ingest.py can reach:
docker run -p 6333:6333 qdrant/qdrant
python ingest.py path/to/paper.pdfDie vollständige Pipeline (Docker-Compose-Stack, /search- und /ingest-API, MCP-Tools) wird funktionieren, sobald main.py existiert.
Ingest a PDF file from the local filesystem into the search index. The file is chunked, embedded (dense + sparse), and stored in Qdrant.
| Name | Required | Description | Default |
|---|---|---|---|
| file_path | Yes | Absolute or relative path to the PDF file. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden of disclosing effects. It clearly explains the processing pipeline: the file is chunked, embedded with dense + sparse vectors, and stored in Qdrant. This lets an agent anticipate side effects. It does not mention idempotence, duplicate handling, or failure behavior, but the core mutation is transparently described.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two tight sentences with no filler. The first sentence states the primary action and target, the second adds the important processing steps. Every word earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a single-parameter ingest operation, the description covers the core workflow and storage destination. It does not describe the return value or failure behavior, but given the simple input and lack of output schema, the definition is reasonably complete for an agent to invoke it correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The input schema already provides 100% coverage for the single parameter file_path, including its type and that it is an absolute or relative path. The description adds little beyond the schema, reinforcing that the file is a local PDF. That meets the baseline of 3 but does not exceed it.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a specific verb ('Ingest'), a clear resource ('PDF file'), and a destination ('search index'). It also names the sibling (search_papers) implicitly by its focus on adding to an index, so an agent can distinguish ingest from search without extra context.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies when to use this tool (when a PDF needs to be added to the search index) versus the sibling search_papers (when querying is needed), but it never explicitly states 'use this when... use search_papers when...' or mentions any exclusions or prerequisites.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Search the ingested paper collection using a natural-language query. Returns the most relevant passages with page numbers and sources, ranked by a cross-encoder reranker.
| Name | Required | Description | Default |
|---|---|---|---|
| limit | No | Number of results to return (default 5). | |
| query | Yes | The search query in natural language. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden of behavioral disclosure. It reveals that results are passages rather than full papers, include page numbers and sources, and are ranked by a cross-encoder reranker, which meaningfully describes behavior beyond the tool name.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is two sentences with the main action front-loaded. The second sentence efficiently packs return format and ranking behavior. No unnecessary words or redundancy.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple two-parameter search tool with no output schema, the description adequately conveys what results include (passages, page numbers, sources, ranked order) and the scope (ingested collection). It does not specify the exact result data structure, but that is minor given the straightforward contract and the clear sibling relationship.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100%, so the baseline is 3. The description reinforces that the query is natural-language but adds no new detail about the limit parameter or query formatting. It does not need to compensate for schema gaps.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a specific verb ('search') and resource ('ingested paper collection'), and clarifies that queries are natural-language. It also distinguishes itself from the sibling 'ingest_paper' by focusing on retrieval of already-ingested content rather than adding papers.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The phrase 'ingested paper collection' clearly implies this is for searching papers previously added, giving the agent a strong contextual signal about when to use it. However, it does not explicitly name the alternative or state exclusions, so it falls short of full routing guidance.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0ingest_papersearch_papersScored across 2 tools
The two tools have completely distinct roles: one adds documents to the index, the other queries it. There is no overlap or ambiguity between ingestion and search.
Both tools follow a verb_noun pattern and are recognizable at a glance. The only minor inconsistency is that one uses the plural 'papers' while the other uses singular 'paper'.
Two tools is a thin surface, falling below the typical 3-15 well-scoped range. Each tool is essential for the core RAG workflow, but the set still feels minimal for anything beyond basic ingest-and-query usage.
The core ingest and search operations are present, giving agents the main RAG loop. However, there is no way to list, delete, or update ingested papers, which leaves obvious workflow gaps around managing the collection.
Search arXiv/Semantic Scholar/OpenAlex + medical evidence (PubMed/Europe PMC) + LaTeX/PDF tools.
Research paper search with real citations and reference formatting for AI assistants
Search arXiv and ACL Anthology, retrieve citations and references, and browse web sources to accel…
Search 340M+ academic papers — citation graphs, semantic similarity, and AI literature reviews.
