The production loop

flowchart LR    A[Source data] --> B[Parse and clean]    B --> C[Chunk and enrich]    C --> D[(Vector index)]    Q[User query] --> R[Retrieve and rerank]    D --> R    R --> G[Grounded generation]    G --> E[Evaluation and traces]    E -. feedback .-> B

The architecture is a loop, not a pipeline. Every answer creates evidence that should improve ingestion, retrieval and evaluation.

Retrieval before generation

For a query $q$ and candidate chunk $d$, a hybrid score can combine semantic and lexical signals:

$$S(q,d)=\alpha S_{dense}(q,d)+(1-\alpha)S_{BM25}(q,d)$$

The exact value of $\alpha$ matters less than measuring performance on a representative evaluation set.

1
2
3
4
5
6

public record RetrievalResult(
    String documentId,
    String content,
    double score,
    Map<String, String> metadata
) {}

What to measure

1. Context recall — did retrieval include the evidence required to answer?
2. Context precision — how much retrieved material was actually relevant?
3. Faithfulness — is the answer supported by the supplied context?
4. End-to-end latency — p50 is pleasant; p95 is reality.
5. Abstention quality — does the system know when it does not know?

Reliability begins when “I don’t have enough evidence” is treated as a valid product outcome.

Start from the access pattern

Consider a query that finds the latest paid orders for one customer:

1
2
3
4
5
6

SELECT id, total_amount, paid_at
FROM orders
WHERE customer_id = ?
  AND status = 'PAID'
ORDER BY paid_at DESC
LIMIT 20;

A useful candidate is:

1
2

CREATE INDEX idx_orders_customer_status_paid
ON orders (customer_id, status, paid_at DESC);

The equality predicates lead, followed by the column that provides ordering. But the index is only a hypothesis until the query plan and production distribution confirm it.

Read the plan as evidence

Pay attention to estimated rows, the chosen access type, extra sorting and whether the optimizer’s assumptions match real cardinality. A clean-looking EXPLAIN can still hide skewed data and p99 pain.

The durable habit is simple: measure the query, understand the access path, change one thing, measure again.

Map the bottlenecks

flowchart TB    A[EDA and IP] --> B[Chip design]    B --> C[Wafer fabrication]    D[Equipment and materials] --> C    C --> E[Advanced packaging]    E --> F[Cloud · Edge · Devices]

The most attractive business is not always the one with the loudest product launch. Long-term research asks where technical difficulty, switching costs, capital intensity and ecosystem lock-in create durable bottlenecks.

Research questions

• Which layer captures value when compute demand grows?
• Where does capacity take years rather than quarters to add?
• Which tools are deeply embedded in customers’ workflows?
• What breaks the thesis: substitution, geopolitics, cyclicality or valuation?

This journal records a research process, not financial advice. A good thesis must include the conditions under which it is wrong.