Skip to main content
Ask Titan Engineering Deep Dive · Part 6 of 7

Enterprise documents and RAG inside Ask Titan

A business question can require a procedure, specification or policy rather than a database calculation. The right document passage must reach the agent first.

An implementation-grounded look at document processing, Qdrant retrieval and the dedicated LangGraph document specialist.

LangGraph MCP Qdrant Embeddings Document retrieval

Ask Titan Engineering Deep Dive

Seven articles, one architecture

Part 6 follows the evidence path from enterprise documents to a specialist answer. The final article examines how recorded execution data helps investigate that answer.

The engineering problem

Not every business answer lives in a table

“How many batches were rejected?” is a structured-data question. “What should we do when a batch fails intake?” may require a quality procedure. An AI assistant needs different execution paths for these two questions.

A document is not a trustworthy answer simply because it has been uploaded. The system must extract usable text, make it retrievable, find relevant passages for this question and let the specialist reason from those passages.

Same factory · different evidence

“How many batches failed intake today?”

Query governed operational data through a structured specialist.

“What does the intake procedure require next?”

Retrieve the relevant instruction from enterprise documents.

Same orchestrator. Different specialist capability.

Ask Titan document indexing and retrieval During indexing, documents are extracted, chunked, enriched with metadata and embedded into Qdrant. At question time, the document specialist prepares a search query, retrieves passages from that same Qdrant collection and uses them to generate an answer. INDEXING · BEFORE THE QUESTION RETRIEVAL · FOR THIS QUESTION SEARCH THE INDEX Documents Policies · specs · guides Prepare Extract · chunk · enrich Embed Vectorize passages Qdrant Searchable collection Question Asked by the user Prepare search Reformulate the query Retrieve Relevant passages Generate answer LLM + retrieved evidence

One Qdrant collection supports both paths: indexing prepares the search surface, retrieval uses it when a question arrives.

Architecture · two connected paths

Index documents. Retrieve when asked.

Ask Titan separates document preparation from question-time reasoning. Documents are processed ahead of time, chunked, enriched and embedded into a Qdrant collection so they can be searched efficiently later.

When a user asks a question, the document specialist does not search raw files directly. It reformulates the request, retrieves the most relevant passages from Qdrant and uses those passages as grounded evidence for the answer.

Step 1 · prepare the evidence

A file becomes searchable passages

Ask Titan sends uploaded files to a separate processing service. It supports PDFs, Word documents, presentations, spreadsheets, CSV, JSON and plain text. The service extracts their content and, where supported, preserves useful structure such as headings, column names and rows.

The extracted content is split into smaller passages and enriched with source metadata and an AI-generated document topic. Each passage is embedded and indexed in Qdrant for later retrieval. Content-derived chunk IDs help avoid duplicate entries during ingestion, but do not automatically manage document versions.

Indexing pipeline

01

Extract

Read text and available document structure from the source format.

02

Split + enrich

Create passages and attach source metadata, content identity and a document topic.

03

Embed + index

Store passage embeddings and associated text in the configured Qdrant collection.

Inside the retrieval boundary

Configured document collection

Qdrant

Passage text · embeddings · metadata

Retrieval result

Relevant passages

Selected document evidence returned to the specialist.

Step 2 · retrieve

The document collection is configured, not guessed by the model

Ask Titan exposes document retrieval through an MCP specialist associated with a configured Qdrant collection. When a user asks a document question, the specialist prepares a search query and retrieves relevant passages from that collection.

The inspected implementation uses vector similarity search. The retrieved passages become evidence for the specialist’s answer; they are not the answer itself. Using Qdrant does not, on its own, imply additional techniques such as hybrid search or reranking.

Step 3 · the document specialist

Three explicit steps. One specialist capability

The top-level orchestrator invokes the document specialist through a permitted MCP tool. Its dedicated LangGraph workflow prepares a standalone search question, retrieves passages from the configured Qdrant collection and passes that evidence to the answer-generating model.

For example, a user asks, “What are the next steps when the intake check fails?” The specialist retrieves relevant procedure passages and uses them to generate an answer. Those passages provide evidence, but they do not automatically verify a generated citation. Part 7 explores how answers can be traced back to their sources and tool execution.

Document specialist workflow The specialist prepares a standalone search question, retrieves passages from its configured Qdrant collection, then uses those passages to generate an answer. 01 · QUESTION REWRITER Prepare the search question 02 · RETRIEVAL Find passages in Qdrant 03 · ANSWER GENERATION Use retrieved evidence

Three separate access boundaries

Capability access

Can this user invoke the MCP tool?

Application policy

Collection scope

Which collection does the specialist search?

Specialist configuration

Document permissions

May this user receive the retrieved passage?

Requires separate enforcement

Tool access ≠ document access

Governance · scope and permissions

Finding a document is not the same as being authorized to read it

Document retrieval crosses different access boundaries. Application policy determines whether a user can invoke the document specialist. The specialist then searches its configured Qdrant collection.

Neither decision automatically establishes permission to read every document in that collection. Document-level access requires separate enforcement before restricted passages can be disclosed.

Engineering takeaway

Document retrieval belongs behind a specialist boundary

The top-level agent should not ingest PDFs or search the entire document corpus itself. Give it a permitted document capability, let that capability select and retrieve passages, then return the evidence-backed response to the conversation.

Patterns we keep

Keep indexing separate from inference.
Preserve source context while extracting.
Bind retrieval to a configured collection.
Rewrite a query before searching when useful.
Treat passages as evidence, not truth.
Verify provenance and access separately.

Next in the series

Part 7 · Answer traceability

Next we connect retrieved evidence, tool execution and user-facing answers. What can an operator or business user actually trace back to its source, and where are additional controls required?

Read Part 7

FAQ

Enterprise document retrieval and RAG questions

Practical answers about documents, chunking, Qdrant retrieval, LangGraph specialists and source boundaries.

What does RAG mean in Ask Titan?

Retrieval-augmented generation means finding relevant passages in enterprise documents and supplying them to the document specialist before it generates an answer. It is separate from querying structured tables.

Which document formats are handled by the inspected upload path?

The repository contains loaders for PDF, TXT, DOCX, PPTX, XLSX/XLS, CSV and JSON. Extraction fidelity depends on the file and loader; scanned-page OCR is not established by this path.

Which vector database does Ask Titan use?

The inspected document path uses Qdrant through a LangChain vector-store integration. Weaviate is a public architecture reference for the illustrations, not the backend used in the inspected implementation.

Does Ask Titan use hybrid keyword and vector search?

The inspected retrieval path uses a standard vector-store retriever. It does not establish a configured BM25 hybrid, reranking or multi-stage search pipeline.

How does the document specialist answer a question?

Its LangGraph workflow prepares a standalone search question, retrieves relevant passages from a configured collection and generates an answer using the retrieved text.

Can the document specialist automatically see every previous chat message?

The inspected specialist tool receives a query and starts its internal graph with that query. Conversation continuity is managed at the top-level orchestrator, not proof that all prior messages reach the specialist.

Are the citations guaranteed to point to the correct source?

The retrieval API returns passage content and metadata, and the specialist instructs the model to cite sources. The inspected answer path does not establish a separate verified citation-mapping mechanism.

Does a permitted document tool automatically enforce per-document access rights?

No. Capability permission, collection choice and individual document access are separate controls. The inspected code is not proof of end-user document-ACL passthrough.

Ask Titan · enterprise knowledge

Put operational knowledge within reach of the people who need it

Procedures and specifications are useful when teams can find the right passage at the right time. Ask Titan connects document retrieval to the same question-driven assistant used for enterprise data.

From source to answer

01Prepare enterprise documents for search
02Retrieve relevant passages for a question
03Generate a response from selected evidence