Citations
Document, page, passage
Every claim carries document, page and passage, so a reader checks it against the original instead of trusting it.
Use case
Ask your own documents. Get the page number.
Policies, manuals, contracts and support archives hold the answers; staff find them by asking whoever remembers. Ask the corpus instead, and get the page.
From a folder of PDFs to answers people check.
Four stages, all inside your infrastructure. Nothing ships out to be embedded, indexed or answered.
01
Convert
PDF, Office, email and scans become clean text with structure intact.
02
Index
Chunks embed locally into the built-in store, or Qdrant and pgvector.
03
Ask
Hybrid retrieval grounds a local model; answers come from passages.
04
Check
Each answer carries document, page and score for one-click verification.
Evidence, not assertion.
An internal search tool survives only if people can verify it. Three properties make that the default.
Refusal
It says when it does not know
When the corpus does not support an answer, that is the answer, stated plainly rather than papered over with confidence.
Scoring
You can see why a result ranked
Retrieval score and strategy per hit, so a poor answer is diagnosed and tuned rather than argued about.
Wording, meaning, or both together.
Wording matters for a part number, meaning matters for a policy question. Run either, or both fused.
Lexical
When the exact term matters
BM25 with language-aware analysis, so part numbers, clause references and product codes match exactly as typed.
Semantic
When the phrasing does not match
Vector retrieval for questions asked in different words from the source, which is most questions about an unread policy.
Hybrid
Fused, with weights you set
Both arms fused by reciprocal rank or normalized score, then reranked; recency, diversity and per-document collapse are request switches.
Less to operate than you would expect.
The stack that usually takes three vendors comes up with the product.
Storage
No database to provision
Full-text and vector search run on a local cluster that ships with the product; PostgreSQL with pgvector and Qdrant are supported too.
Boundary
Embeddings stay put too
Corpus, chunks and vectors are all local. Nothing ships out to be embedded, the step that quietly breaks residency requirements.
Assistants
Controlled access for agents
An external assistant queries the corpus through a governed MCP tool and receives ranked passages, never the filesystem or the whole collection.
Assistants working this corpus through MCP have their own use case.
Load, ask, cite.
Load documents, ask in natural language, get grounded answers with page citations. The server exposes the same over HTTP and MCP.
using LMKit.Model;
using LMKit.Retrieval;
var chatModel = LM.LoadFromModelID("qwen3.5:9b");
var embedModel = LM.LoadFromModelID("embeddinggemma-300m");
using var chat = new PdfChat(chatModel, embedModel);
await chat.LoadDocumentAsync("employee-handbook.pdf");
await chat.LoadDocumentAsync("it-security-policy.pdf");
await chat.LoadDocumentAsync("data-retention-policy.pdf");
var answer = await chat.SubmitAsync(
"How long do we retain applicant data?");
Console.WriteLine(answer.Response.Completion);
// Grounded answer with passage retrieval and page citations.# The same grounded answer from LM-Kit One's search service.
curl http://your-server:5189/lmkit/v1/search/answer \
-H "Authorization: Bearer your-server-token" \
-H "Content-Type: application/json" \
-d '{"clusterId":"internal","tenantId":"<org-tenant-id>",
"collectionId":"<policies-id>",
"question":"How long do we retain applicant data?"}'
# answer + citations[] + groundednessScore + unsupportedClaimsThe high-level document conversation: a session that holds the loaded document across questions, with follow-ups, history control, and clean teardown.
curl http://your-server:5189/lmkit/v1/chat-with-document \
-H "Authorization: Bearer your-server-token" \
-H "Content-Type: application/json" \
-d '{"sessionId":"policy-review",
"input":"<file id from /lmkit/v1/files>",
"inputFormat":"FileIdentifier",
"question":"How long do we retain applicant data?"}'
# Follow-up questions reuse the session; the document stays loaded.The technology underneath.
The corpus pipeline composes documented capabilities you can inspect on their own.
Answer
RAG chat
Conversational retrieval with citations, four query modes, streaming.
How RAG chat worksStore
Built-in vector database
Embedded storage that ships with the product; Qdrant and pgvector supported.
How storage worksConvert
Document to Markdown
PDF, Office, email, and scans become clean, structured, indexable text.
How conversion worksThen scale
From one corpus to the backbone
When several teams and tools need the same retrieval, LM-Kit One's search service turns this pipeline into a governed, multi-collection backend on the database you already run.
The enterprise search use caseFrequently asked questions.
How do we know an answer is right?
You check it, quickly. Every answer cites document, page, and passage with its retrieval score, and when the corpus does not support an answer, the system says so instead of improvising. Trust comes from verifiability, not from tone.
What formats can the corpus contain?
PDF, Office documents, email, HTML, Markdown, and scans. Conversion and OCR normalize everything to structured text before indexing, so a scanned policy answers questions as well as a born-digital one.
Where do the embeddings live?
On your infrastructure, next to the documents. Nothing ships out to be embedded or indexed, which is the step that quietly breaks data-residency requirements in hosted knowledge tools.
What is the path from one corpus to company-wide search?
The same engine scales two ways: keep it embedded with LM-Kit.NET for a single application, or deploy LM-Kit One and let its search service host per-team collections with grants, audit, and horizontal scaling.
Private Document Intelligence