Document Intelligence

Document intelligence that actually reads your business.

Drop your contracts, policies, and reference material in once — then your chat, your skills, and your customer-facing assistants answer from them, with citations you can click. No re-uploading. No copy-paste. And when the answer isn’t in your documents, it says so instead of guessing.

01 · What it is

Retrieval-augmented generation, done properly.

“RAG” means the assistant retrieves the most relevant passages from your own documents before it answers — so responses are grounded in your material, not the model’s general training. Plenty of tools claim RAG. The difference is in the retrieval quality and the discipline around citations. This page walks through both — an accessible overview first, then the full technical pipeline for the evaluators who want to see the engineering.

02 · Why it’s different

Built to be trusted with the documents that matter.

Grounded and cited

Answers quote your documents with inline citations — and say “I couldn’t find that in your documents” instead of inventing an answer. No confident hallucinations.

Catches exact references

Vector-only search misses precise tokens like a statute or policy number. Our hybrid retrieval pairs meaning with exact keyword matching, so “IC 6-1.1” surfaces every time.

Isolated per workspace

Retrieval runs under the same database-level row security as the rest of the platform. Your documents can only ever answer your questions.

Reads real-world files

PDFs, Word docs, spreadsheets and text up to 100 MB — with page-accurate citations that open a PDF to the exact page.

03 · The technical deep-dive

How it works, end to end.

From the moment a document is uploaded to the moment a cited answer comes back, the system runs six stages. Each one exists to fix a specific way naive RAG goes wrong — chunks that cut sentences in half, vector search that misses exact numbers, or a model that confidently makes things up.

04 · The pipeline

Six stages, one cited answer.

01 · IngestDocuments become clean, page-tracked text
PDFs (with per-page tracking), Word, and text/CSV are extracted in the background. Large files upload directly to cloud storage, bypassing request-size limits, and a self-healing sweep retries anything that stalls.
02 · ChunkSplit on natural boundaries
Text is divided into ~2,000-character chunks with ~200 characters of overlap. Breaks prefer paragraph, then sentence, then word boundaries — so a chunk is a coherent passage, not a sentence cut in half.
03 · EmbedMeaning becomes vectors
Each chunk is embedded with OpenAI’s text-embedding-3-small (1536 dimensions) and stored in pgvector with an HNSW index for fast cosine-similarity search.
04 · Retrieve (hybrid)Semantic + keyword, fused
A query runs two searches at once — vector similarity for meaning and Postgres full-text for exact terms — then fuses both ranked lists with Reciprocal Rank Fusion. Exact keyword hits bypass the similarity floor, so precise references never get filtered out.
05 · RerankA model re-sorts for true relevance
The fused candidates are re-ranked by a fast Claude Haiku pass that reads the passages against the question. In legal mode it uses a statute-aware rubric that favors the controlling section. It’s best-effort — if it ever fails, retrieval order stands; answers never break.
06 · GroundTop passages, with citations
The best passages are handed to the assistant tagged like [doc:filename#chunk-3 p.5], with an instruction to cite what it uses and admit when the answer isn’t there.
05 · Grounding & isolation

Citations you can verify, isolation you can trust.

Retrieval quality is only half the job. The other half is making the answer accountable — and keeping every workspace’s documents apart.

Cite or admit

The assistant is instructed to ground its answer in the retrieved passages, cite the ones it relies on, and say it couldn’t find the answer rather than invent one. The same discipline applies in chat, skills, and Document Search.

Open to the exact page

For PDFs, each chunk carries its page number, so a citation can open the source document at the precise page the claim came from.

Per-workspace by construction

Every retrieval runs inside your workspace’s database context, and the chunk store is governed by strict row-level security. Public assistants are further fenced off from internal admin and governance documents. See the isolation model →

06 · Recently shipped & what’s next

The engine keeps getting sharper.

Document intelligence is an area we’re actively investing in — most recently with an attorney-grade retrieval upgrade.

v1.0.83
Document RAG launches — semantic retrieval and a dedicated Document Search.
v1.0.86
Page-accurate citations — open a PDF straight to the cited page.
v1.0.98–100
Reliability — bulk upload, larger-file handling, auto-heal, and embedding rate-limit pacing.
v1.0.103
Hybrid search + reranking + grounded answers — the attorney-grade upgrade.

On the roadmap

Jurisdiction-aware retrieval — scope answers to a chosen body of law.
Structured legal source types — distinct handling for statutes, regulations, and case law.
Query expansion — match common synonyms and citation formats automatically.

Roadmap items are in development and not yet generally available.

Put your documents to work

Let the assistant read your business.

Upload your documents once and get cited, grounded answers across chat, skills, and customer-facing assistants. Start with a free assessment.