01 / 05 · Research
Papers that answer back.
Query across literature libraries, systematic reviews, and supplementary appendices with precision that "nearest neighbour" can't even imagine.
- § Abstract
- § Methods
- § Results
- § Discussion
- § References
Vectorless SDK · v1.0 · early access
Retrieval, withoutthe vectors.For agents that read.
Vectorless is the retrieval primitive for AI agents. We turn your PDFs, docs, and knowledge bases into structured maps any LLM can navigate — no chunking, no embeddings, no vector DB to operate.
No credit card · MIT licensed · 100 documents free
$npm i @vectorless/sdk
Scroll the story
avg retrieval
~40ms
parallel section fetch
embedding cost
$0
we never compute one
chunks tuned
0
no overlap, no regret
works with
any LLM
OpenAI, Anthropic, Gemini, local
Fits in your stack
◎OpenAI✦Anthropic◇Gemini⌖LangGraph🦙LlamaIndex◈CrewAI↯MCP⌘Cursor▲Vercel
The indictment
Mostretrievalstartsbydestroyingyourdocument.
Chunks lose context. Embeddings miss meaning. Similarity isn't comprehension. We've been retrieving with a blindfold on for years.
research.pdf
page 12 / 42
chunked into oblivion
The crossroads
One starts with vectors. The other starts with sense.
Vectorful.
Slice the document. Embed every fragment. Hope the math finds meaning.
verdict
Fragments fly. Context dies.
Vectorless.
Keep the document whole. Map its structure. Let reasoning navigate.
01Abstract
02Introduction
03Methods
04Results
05Discussion
06Conclusion
verdict
Structure stays. Meaning lands.
Lessmath.Moremeaning.
The method
Three steps down. No vectors required.
Step 01
Ingest.
Add your document.
Drop in a PDF, DOCX, URL, or raw text. We read the whole thing — title, headings, footnotes, the lot — and store every section as a stable, addressable unit. No chunk size to tune. No overlap to tweak.
“No knife required.”
01
Step 02
Map.
Get a navigable table of contents.
You get back a structured manifest — every section with a title, a summary, and a deterministic link. Think llms.txt, but generated for you, on anything. A document, finally treated like a real interface.
“A map your LLM can actually read.”
02
Step 03
Reason.
Let the LLM choose what to fetch.
Hand the map and a question to any model. It reasons about which sections matter, we fetch them in parallel, and you get back complete, structured context — not the nearest neighbour, the right answer.
“Retrieval that knows where it's going.”
03
In code
Five lines. Every retrieval problem.
Bring your own LLM. Bring your own framework. We just hand you a primitive that fits where vectors don't.
// 01 — Ingest
const { doc_id, toc } = await vectorless.addDocument(file);
// 02 — Map
const sectionIds = await llm.reason(toc, question);
// 03 — Reason
const context = await vectorless.fetchSections(doc_id, sectionIds);
const answer = await llm.answer(context, question);The anatomy
What your documentbecomes inside Vectorless.
Not a wall of vectors. A navigable map. Every section gets a stable address, a title, a summary, and a deterministic link any LLM can call.
research.pdf
42 sections · 187 pages
§01 · Abstract
§02 · Introduction
§03 · Methods
§04 · Results
§05 · Discussion
§06 · Conclusion
the source
toc.manifest
json- section01titleAbstractlink/doc/0x7f/§01
- section02titleIntroductionlink/doc/0x7f/§02
- section03titleMethodslink/doc/0x7f/§03
- section04titleResultslink/doc/0x7f/§04
- section05titleDiscussionlink/doc/0x7f/§05
- section06titleConclusionlink/doc/0x7f/§06
fetch in parallel~40ms · any LLM
the interface
Everything retrievalshould have been.
→ doc01
Structured ingest
PDFs, DOCX, TXT, URLs, raw text. We pull every section out as an addressable unit — automatically. No tuning, no overlap, no regret.
manifest02
LLM-navigable maps
Every document becomes an llms.txt-style manifest — a map any model can read natively. Built so machines can think, not search.
|| fanout03
Parallel section fetching
Fetch one section or twenty. Latency is the same. Fan-out retrieval without the fan-out tax.
⇄ hybrid04
Hybrid retrieval
Not every document is structured. Toggle embedding-based fallback alongside reasoning when you need it. We won't judge you for keeping a few vectors around.
§ link05
Deterministic links
Every section gets a stable URL. Retrieval is auditable, reproducible, and grep-able. Citations come for free.
⌘ open06
Any LLM, any framework
Vectorless is a primitive, not a platform lock-in. Drop it into LangGraph, CrewAI, LlamaIndex, an MCP server, or your own stack. We retrieve. You decide.
✎ toc07
Auto Table of Contents
No headings? We generate a semantic ToC with an LLM — structure from the unstructured. Even your messiest doc gets a spine.
∎ pg08
Postgres native
JSONB for maps, pgvector for fallback, full-text built in. One database. No vector infra to babysit, no separate cluster to forget about.
The audiences
Built for documents that actually matter.
02 / 05 · Legal
Contracts that hold their shape.
Clauses are units. Sub-clauses are units. Vectorless retrieves whole, intact context — no clause torn in half by an arbitrary chunk size.
- § Definitions
- § Term
- § Indemnity
- § Termination
- § Governing Law
03 / 05 · Clinical
Guidelines without guesswork.
Dosing tables, contraindications, decision trees — retrieved as complete sections, not bisected fragments. Patient safety, not similarity scores.
- § Indications
- § Dosing
- § Contraindications
- § Warnings
- § Adverse Effects
04 / 05 · Technical docs
APIs your agent can actually read.
Endpoints, parameters, error codes — structured the way developers already think, addressable the way LLMs already navigate.
- § Auth
- § Endpoints
- § Parameters
- § Errors
- § Webhooks
05 / 05 · Enterprise
Wikis with working memory.
Internal SOPs, onboarding handbooks, runbooks — every section addressable, every retrieval auditable, every answer traceable.
- § Onboarding
- § Security
- § Engineering
- § HR
- § Finance
preprint.pdf
§ 187 pp- 01Abstract
- 02Methods
- 03Results
- 04Discussion
- 05References
GET /doc/research/§04200 OK
MSA-v3.docx
§ clauses- 01Definitions
- 02Term
- 03Indemnity
- 04Termination
- 05Governing Law
GET /doc/legal/§04200 OK
protocol-2026.pdf
rx- 01Indications
- 02Dosing
- 03Contraindications
- 04Warnings
- 05Adverse Effects
GET /doc/clinical/§04200 OK
reference.md
api- 01Auth
- 02Endpoints
- 03Parameters
- 04Errors
- 05Webhooks
GET /doc/technical docs/§04200 OK
handbook.notion
kb- 01Onboarding
- 02Security
- 03Engineering
- 04HR
- 05Finance
GET /doc/enterprise/§04200 OK
Built by people who got tiredof tuning chunkers.
I spent six months tuning a chunker so the retrieval would stop bisecting our clauses. Vectorless made that whole problem disappear in an afternoon.
Eze Onyekpere
Staff Engineer, Legal AI
Reasoning over a table of contents is the obvious right answer in hindsight. I don't miss managing a vector DB.
Priya Sundaram
AI Lead, Clinical Research
Our agent's answers got noticeably better the day we swapped in Vectorless. Same prompts, same model, structured retrieval.
Marcus Le
Founder, dev-tools startup
The SDK
A retrieval primitivethat fits anywhere.
TypeScript first, Python coming. The surface is small on purpose — ingest and retrieve. Your reasoning, your LLM, your orchestration. We just handle structure.
- addDocumentput
- getToCread
- fetchSectionpick
- fetchSectionspick·N
- listDocumentslist
↯
MCP server coming — call Vectorless natively inside Claude, Cursor, and any MCP-compatible runtime.
vectorless.addDocument(file, options)
// → { doc_id, toc }
vectorless.getToC(doc_id)
// → ToC manifest
vectorless.fetchSection(doc_id, section_id)
// → { title, content, page_range }
vectorless.fetchSections(doc_id, section_ids)
// → Section[]
vectorless.listDocuments()
// → DocumentSummary[]Common questions
Honest answersto the obviousquestions.
Can't find your question? Open an issue on GitHub or email hello@vectorless.dev.
The opposite — we removed the steps. No chunker, no embedder, no vector DB, no similarity threshold. The LLM reads a structured table of contents and asks for the sections it needs by ID. Retrieval becomes a function call, not a math problem.
Yes. If a document doesn't come pre-structured, we generate a semantic table of contents with an LLM during ingest. You get the same navigable map — auto-built — for messy PDFs, scanned reports, and plain text.
You skip the embedding bill entirely (no per-token embedding fees, no upserts). Storage is plain Postgres JSONB plus the raw document. The only variable cost is the LLM reading the ToC at query time — usually ~200 tokens.
The ToC is hierarchical. For a 500-page legal contract, the LLM first picks a top-level section, then drills down. Two reasoning hops, one fetch — works the same way a human would navigate.
Yes — hybrid retrieval is supported. Toggle embedding fallback for sections that don't map cleanly, or run them in parallel. We're not religious about it, we just don't make you start there.
Any. Vectorless ships as a TypeScript SDK (Python coming) and is intentionally narrow — ingest and retrieve. Use it inside LangGraph, CrewAI, LlamaIndex, an MCP server, or your own loop. Bring your own model.
Yes — the SDK is open source and free for development. The hosted plan is free for the first 100 documents; paid plans start when you scale beyond that.
Same instinct, more surface area. llms.txt is a static convention for one site; Vectorless generates a queryable manifest for any document, makes every section addressable, and handles the fetch.
Pricing
Free to start.Honest as you scale.
Hobby
Free, forever
$0
free up to 100 docs
Everything you need to ship a real agent.
- 100 documents
- Unlimited section fetches
- TypeScript + Python SDK
- Hybrid retrieval (BYO embeddings)
- Community support
Most popular
Pro
For shipping teams
$29
per seat / month
Scale past prototype without paging an SRE.
- 10,000 documents
- Priority parallel fetch
- MCP server integration
- Document analytics + audit log
- Email support · 24h SLA
Enterprise
For regulated work
Custom
talk to us
SSO, on-prem, contracts, and people who pick up the phone.
- Unlimited documents
- SSO + SCIM
- Self-hosted / VPC
- SOC 2 + DPA
- Dedicated Slack channel
The parting line
Stopsearching.
Startreasoning.
Free for 100 documents. No credit card. Five lines of code from install to your first answer.
no chunks·no embeddings·no regrets·structure first·reasoning second·retrieval, solved