Dvcon papers

DVCon Papers

This skill gives the agent direct access to the local DVCon paper corpus through the dvcon MCP server (defined in backend/src/backend/mcp_server.py). The corpus is the same one served by the FastAPI backend and React UI; the MCP server just re-exposes it as tools so agents can query it without going over HTTP.

When to use

Trigger this skill when the user wants to:

• search the DVCon proceedings by keyword, topic, author, method, or concept
• read a specific paper's extracted markdown or full metadata
• summarize one paper or compare several papers
• look at the author / conference / company / reference graph for a paper
• ask a grounded question whose answer must come from the indexed papers
• get corpus counts (how many papers, years, locations)

Prerequisites

The dvcon MCP server must be configured in the client's MCP settings and pointed at uv run --project backend dvcon-mcp (stdio). The corpus lives under data/ and is configured via .env. The read tools (search, detail, markdown, graph, stats) work without GROBID or OpenAI. The chat_with_papers tool requires OPENAI_BASE_URL and OPENAI_API_KEY.

If the corpus is empty, seed it first with uv run --project backend ingest --limit 5 (or the /api/admin/ingest endpoint), then re-run search.

Available tools

Tool	Purpose
search_papers	Keyword / semantic / hybrid search with year + location filters
get_paper_detail	Full metadata for one paper (abstract, authors, affiliations, references)
get_paper_markdown	Extracted markdown body of a paper (includes image refs)
get_paper_graph	Cytoscape-style nodes/edges for paper relationships
corpus_stats	Paper count, years, locations, conference count
chat_with_papers	Grounded Q&A; constrain to selected paper ids when comparing

Workflows

1. Find papers on a topic

Call search_papers with a free-text query. Hybrid mode (default) merges SQLite FTS5 keyword hits with bge-m3 semantic hits. Example call:

search_papers(query="UVM register abstraction layer", mode="hybrid", limit=10)

Expected return shape:

{
  "mode": "hybrid",
  "count": 3,
  "items": [
    {
      "paper_id": 7,
      "title": "...",
      "abstract": "...",
      "authors": ["..."],
      "affiliations": ["..."],
      "year": 2024,
      "location": "united states",
      "conference_name": "DVCon United States 2024",
      "score": 0.82,
      "snippet": "..."
    }
  ]
}

Leave query empty to browse. Filter with year=2024 or location="india".

2. Read a paper

After finding a paper_id, fetch its detail and markdown:

get_paper_detail(paper_id=7)
get_paper_markdown(paper_id=7)

The markdown response includes the raw extracted text with markdown-relative image references (images/...). Quote specific sections when answering the user.

3. Summarize or compare papers

Pass one or more paper_ids to chat_with_papers to constrain the answer to those papers. With a selection, the server sends full paper text when it fits the model context window, else curated sections. Without a selection, retrieval runs corpus-wide.

chat_with_papers(
  question="Compare the formal verification approaches in these two papers.",
  selected_paper_ids=[7, 12]
)

Returns:

{
  "answer": "...",
  "citations": [{"index": "1", "paper_id": "7", "title": "...", "year": "2024"}],
  "scope_paper_ids": [7, 12]
}

Citation index numbers correspond to the order of selected_paper_ids / retrieval.

4. Explore relationships

get_paper_graph(paper_id=7)

Returns nodes and edges for the paper's conference, authors, companies, and references. Node type is one of paper, conference, author, company, reference.

Output contract

• Always cite paper_id (and title + year when space allows) when referencing a paper.
• When the user asks to "summarize this paper" without a selection, first run search_papers to disambiguate, then chat_with_papers with the resolved id.
• If a tool returns {"error": "..."}, surface the error rather than fabricating a paper that does not exist in the corpus.
• Never invent DVCon paper content that did not come back from a tool call.

Troubleshooting

Symptom	Fix
search_papers returns empty for every query	Corpus is empty — run uv run --project backend ingest --limit 5
chat_with_papers returns an error about OpenAI config	Set OPENAI_BASE_URL and OPENAI_API_KEY in .env
Semantic results look stale after an embedding-model change	Force a reindex: uv run --project backend ingest --limit 1 --force
get_paper_markdown says file is missing on disk	Re-ingest the paper; the markdown tree may have been cleared
MCP tools are not available	Ensure the dvcon MCP server is enabled in the client's MCP settings