From a search question to useful sources

The animation explains the architecture; no live actions are executed.

This overview also includes design choices for further development.

1 Understand the question

• Intent · entity resolution · date and time · source selection · hard constraints

I establish intent, identity, period and permitted sources before selection.

Exact / FTS

• Names · numbers
• Exact terms

Exact lookups and full-text search find names, numbers and literal terms.

Vector

• Semantic
• similarity

Vector search finds passages with related meaning. Identity needs additional evidence.

Structured

• Sections · tables
• Content objects

Sections, tables and content objects retain headings, context and units.

Graph

• Relationships · aliases
• Entity expansion

Supported relationships can lead to other relevant sources. Search depth and relationship types are bounded.

Memory

• Preferences
• Earlier corrections

Recorded preferences and corrections help with selection.

Apps / web

• Email · calendar
• Current sources

The appropriate app or external source is queried, such as the calendar for an appointment.

3 Candidate fusion

• Normalise candidates
• Combine sources

Candidates are combined while preserving source identity and retrieval route.

4 Reranking

• Relevance · identity · time
• Authority · personal context

Ranking weighs relevance, identity, time, quality and provenance for this question.

7 Context package

• Passages + relationships + source
• Compact, preserving context

Selected passages and necessary context go into a bounded package.

8 Enough evidence?

• Right person, period and source?
• Are all subquestions covered?

I check which parts of the question the sources support and what is missing.

Answer / preparation

• With source location and meaning
• Focused on the user question

The answer cites sources and distinguishes facts, inferences and proposals.

9 Targeted follow-up search

• Use new clues, don’t repeat
• Bounded passes + budget

A missing piece can lead to a targeted follow-up query within time and step limits.

Connections

• 1 Understand the question → Exact / FTS
• 1 Understand the question → Vector
• 1 Understand the question → Structured
• 1 Understand the question → Graph
• 1 Understand the question → Memory
• 1 Understand the question → Apps / web
• Exact / FTS → 3 Candidate fusion
• Vector → 3 Candidate fusion
• Structured → 3 Candidate fusion
• Graph → 3 Candidate fusion
• Memory → 3 Candidate fusion
• Apps / web → 3 Candidate fusion
• 3 Candidate fusion → 4 Reranking
• 4 Reranking → 7 Context package
• 7 Context package → 8 Enough evidence?
• 8 Enough evidence? → Yes → Answer / preparation
• 8 Enough evidence? → Missing evidence → 9 Targeted follow-up search
• 9 Targeted follow-up search → Targeted new query → 1 Understand the question

How to read this

1. I establish intent, identity, period and permitted sources before selection.
2. Exact lookups and full-text search find names, numbers and literal terms. Vector search finds passages with related meaning. Identity needs additional evidence. Sections, tables and content objects retain headings, context and units. Supported relationships can lead to other relevant sources. Search depth and relationship types are bounded. Recorded preferences and corrections help with selection. The appropriate app or external source is queried, such as the calendar for an appointment.
3. Candidates are combined while preserving source identity and retrieval route.
4. Ranking weighs relevance, identity, time, quality and provenance for this question.
5. Selected passages and necessary context go into a bounded package.
6. I check which parts of the question the sources support and what is missing.
7. The answer cites sources and distinguishes facts, inferences and proposals.
8. A missing piece can lead to a targeted follow-up query within time and step limits.