Akash Goyal

Posted on Oct 4

Paper Lineage: every AI idea has ancestors (built on Sanity Context, Knowledge Bases, App SDK and Workflows)

#devchallenge #sanitychallenge #sanity #ai

Sanity Challenge Path Two Submission

This is a submission for the Sanity Challenge, Path Two: Vibe-Code Something Strange

What I Built

Paper Lineage answers "where did this AI idea come from?" with a chain you can check. It's for anyone reading AI research who wants to know what a paper really built on: students, researchers, engineers doing a literature review.

The problem. Ask an AI tool "where did BLIP-2's Q-Former come from?" and you get a confident paragraph: "it extends Flamingo's resampler…". Citation graphs say that one paper cites another, not how it built on it. Language models fill the gap with plausible verbs, and nothing tells you which verb is wrong.

The app. An evidence-backed family tree of 197 AI papers, traced three generations back from BLIP-2: 1,349 citation links and 205 concepts. It keeps two kinds of claim apart:

a fact: "BLIP-2 cites Flamingo 5 times, here are the sentences", mined from the paper's full text
an interpretation: "BLIP-2 extends BLIP", shown only after a curator has accepted it

Bahdanau attention (2014) → Transformer (2017) → Set Transformer (2018) → Perceiver (2021) → Flamingo (2022) → BLIP-2 (2023)
Perceiver: "Most closely related to our work is the Set Transformer."
Flamingo: "…a Perceiver-based architecture that can produce a small fixed number of visual tokens…"

The Q-Former's ancestry as Paper Lineage shows it. Every hop is a citation, with the sentence the later paper used.

Architecture

Solid red: Sanity features. Peach: our apps built on Sanity. Blue: our code. Grey: third party.

A visitor asks: the Next.js site → our /api/ask agent → Sanity Context MCP (GROQ over the graph, plus a Knowledge Base) → every card re-read from the Content Lake.
A curator reviews: the Lineage Desk (App SDK) accepts a link → Content Lake → Live Content API → the site shows the relation at once.
A new paper arrives: Sanity Workflows runs paper-intake (fetching → checks → curation → publishing).

Six features, each built on Sanity

Feature	Sanity product	Advantage
Ask: a chat that answers with evidence	Sanity Context (MCP, GROQ mode)	The agent writes GROQ over typed references; no vector database or RAG pipeline to build
Paper explanations in Ask	Knowledge Base (Context MCP, Knowledge Base mode)	40 papers uploaded, Sanity builds and serves 20 entries
The evidence graph	Content Lake + GROQ + private id paths	One query walks a chain; full texts and visitor questions stay private on a public dataset
Live verified labels	Live Content API	An accepted link appears on the site at once: no redeploy, no websocket code
Lineage Desk for curators	App SDK (+ Agent Actions)	Real-time document hooks, accept-and-publish in one transaction, Sanity login
Paper intake	Workflows	The engine refuses approval while any link is unreviewed, not the UI

Plus Studio for schema-driven editing. Everything runs on Sanity's free plan.

Demo

Live: https://paper-lineage-xi.vercel.app

Try asking: "Where did BLIP-2's Q-Former come from?" · "What did CLIP build on?" · "How does the Q-Former use learnable queries?"

Ask: one answer, four Sanity sources.

The lead paragraph: written by the model from what it retrieved through Sanity Context MCP: the Q-Former's concept record and its buildsOn chain, via GROQ.
The chain card (BLIP → BLIP-2, "cites 9× · not yet reviewed"): an influence document found with Context MCP (GROQ mode), then re-read from the Content Lake by our server. It says "cites", not "extends", because the GROQ returns a relation only for curator-accepted links.
The paper cards ("built on 16 · 0 built on it"): counts from the Content Lake, rebuilt by id on the server, never typed by the model.
The quote: from the Knowledge Base (Context MCP, Knowledge Base mode), checked word for word against the entry and captioned as a Knowledge Base summary, not the paper's words.

The concept page: the Q-Former's ancestry, six steps back to 2014.

The steps: concept documents linked by curated buildsOn references in the Content Lake, read with GROQ (the concept with its chain, then the links between its papers).
The quotes and "cited 5× · §Model Architecture": the citation fact on each influence document, mined from the later paper's full text.
"Cites · not yet reviewed": the link's review status in the Content Lake; it changes when a curator accepts it in the Lineage Desk.

Explore: BLIP-2's family tree, CLIP selected.

The graph: paper and influence documents from the Content Lake via GROQ: 197 papers, 1,349 links, here trimmed to 30 in year columns.
"27 papers it built on · 22 built on it": counted from the influence links GROQ returns (rejected links excluded).
"Verified only": filters on the review status curators set in the Lineage Desk (App SDK).

Status: N of 1,349 links verified so far. Curation is ongoing; every unverified link shows as "cites ×n", by design.

Code

https://github.com/akashgoyal/paper-lineage

web/ (Next.js site + Ask agent) · studio/ · desk/ (App SDK) · workflows/ (definition + runner) · scripts/ (harvest → curate → seed). The full build log is docs/BUILD_LOG.md.

My Build Process

Two days (Oct 2–3) in Claude Code, from an empty folder to the deployed site, Studio, Desk, workflow and Knowledge Base. I wrote the prompts; Claude Code wrote the code and docs.

Prompts that worked

"create a new repo - start with system architecture - make most use of sanity features which are available in free version… Then create a spec before actual implementation." Architecture and spec came first, so later prompts could be one line.
"Review the design spec. Is it complete… Can it be used for moving ahead with implementation": it found the chat had no way to choose what to show, which led to server-verified display tools.
"Can you do sanity checks with some API code - rather than login in claude browser?": every Studio and Desk query and write path is tested against the real dataset.

Where I had to course-correct

"If paper apis are not working… download the pdf & analyse it on your own." Semantic Scholar returned 429 on every call. Reading full text gave a better signal anyway: where and how often a paper is cited.
"In 205 papers, there are only 48 concepts." The taxonomy became 25 themes and 180 concepts.
"The first page should be a chat-box." The design had been browse-first.
"Highlight clearly in the block diagram the features used are from sanity or code." That gave the red Sanity block in the diagrams.

Where the model got it wrong, and how it was caught

Paid-only features. It planned Comments, Tasks and Releases. Caught when it checked Sanity's pricing page.
"Ancestors" newer than their descendants. Validation found 19 links where the earlier paper was published after the citing one (a later revision cited newer work). Fixed in the pipeline; the 19 remain as unreviewed, for a curator to reject.
Browser reads failing on the live site. The domain wasn't in the project's CORS origins. Caught by screenshotting the deployed site.
The chat model. My credits were on Together AI. I tested the cheapest models through the real agent loop: Qwen3.5-9B returned empty answers, GLM-5.3-Flash never stopped searching, gpt-oss-120b printed its reasoning as the answer. DeepSeek-V4-Flash worked, at about 1¢ per answer.
Model behaviour, fixed in code, not the prompt. It narrated ("Let me show…"), rewrote its lead every turn, produced too many cards, and once drew a chain through a link that doesn't exist. Now there's a card budget, a narration filter, one final lead, and chains need a link at every hop.

Reaching past the Studio

Sanity Context and the Knowledge Base. The docs I had pointed to the Dashboard, but the current CLI has sanity context: the Knowledge Base was created, imported (42 sources, 206 MB) and built from the terminal. The two MCP endpoints still needed the Dashboard. Two surprises: one endpoint serves one mode (a dataset source silently outranks a Knowledge Base on the same endpoint), hence two endpoints. And Knowledge Base entries are Sanity's summaries, not the papers' words, so quotes from them kept failing my word-for-word check. They're now captioned "Knowledge Base summary of BLIP-2".

App SDK: the Lineage Desk. A curator reads the evidence beside each link and accepts or rejects it. That's one transaction with Undo, and the Live Content API updates the public site. The Desk also has an Inbox of questions visitors couldn't get answered, and a story composer that only uses accepted links. It's deployed into the Sanity Dashboard, so curators sign in with Sanity.

Workflows: paper intake. The model first modelled "workflows" as a status field. Workflows is a real early-access engine, so the design was rebuilt around a deployed definition. The useful part: approval is gated in the definition. While any link into the paper is unreviewed, the engine refuses Approve, and the Desk shows the engine's own reason. Sanity Functions schedules run daily on the free plan and hosted workflow runtimes aren't available yet, so a small runner executes each stage's effect (arXiv fetch, link count, publish).

Cut: AI-proposed links (links come from the full-text harvest), a path-finder page, live cursors in the composer.

Not verified by the agent: clicking through the Desk and Studio, which need a Sanity login. Their queries and write paths were tested by scripts against the real dataset.

Sanity Project Details

Project ID: jd22zcim. Dataset production is public: first five papers
Schema: paper, influence (fact / interpretation / review), concept (theme → concept, buildsOn), storyline, gap, plus private paperText.* and questions.*
Context: organisation o2qzzix4g; MCP endpoints paper-lineage-graph (GROQ mode) and paper-lineage-papers (Knowledge Base kbawj3190IH1)
Studio (curators): https://paper-lineage.sanity.studio · Desk: an App SDK app in the Sanity Dashboard

Where facts and interpretation live. One real influence document, trimmed:

// BLIP → BLIP-2
{
  from: {_ref: "paper-2201-12086"},          // BLIP
  to:   {_ref: "paper-2301-12597"},          // BLIP-2
  citation: {                                 // FACT: read-only, mined from BLIP-2's full text
    mentions: 9, methodMentions: 3,
    contexts: [ /* 4 sentences: Introduction, Related Work, Method ×2 */
      {_key: "c2", section: "Method",
       text: "Inspired by BLIP (Li et al. 2022), we jointly optimize three pre-training objectives…"} ]
  },
  evidenceKey: "c2",                          // INTERPRETATION: the sentence that best supports…
  relation: "extends",                        // …this relation. Public only once accepted.
  provenance: {origin: "harvest", suggestedRelation: "extends", confidence: 0.95, reviewDecision: "proposed"}
}

Public queries read interpretation only through select(provenance.reviewDecision == "accepted" => relation), and a unit test fails if any query skips that gate.

How Ask uses Sanity Context.

One real run of "Where did BLIP-2's Q-Former come from?", every call in order:

→ groq_query          find BLIP-2
→ groq_query          concept "Q-Former" → introducedBy, buildsOn[] → introducedBy
→ knowledge_base_read entry "blip2"
→ groq_query          influence links into BLIP-2 from Flamingo and BLIP, with
                      "relation": select(provenance.reviewDecision=="accepted" => relation)
→ showChain × 2       our server rebuilds each card from the Content Lake and checks it
→ suggestFollowUps

3 model turns, 4 Sanity lookups, 23 s, 0.6¢. The model never writes a card: it passes ids, and the server rebuilds the card from Sanity, refusing unknown ids, papers out of date order, chains with a missing hop, inexact quotes and relations on unreviewed links. Each question is saved under a private id path, and a weak answer becomes a gap in the Desk Inbox.

Agent Session

(link to the exported Claude Code session, made public)
Agent Session - 1

Agent Session - 2

DEV Community