Alternative A: Keep the current PostgreSQL full‑text

Flow

raw chunkText
  -> plainto_tsquery
  -> search_tsv @@ query
  -> ts_rank_cd
  -> return top JSONL

Advantages

It is already implemented.
It can be done with PostgreSQL alone.
The infrastructure is simple.
It matches directly with tables that already have search_tsv.

Disadvantages

It is not BM25.
It is weak with long code chunks.
Conditions created by plainto_tsquery can become overly narrow.
There is a lot of code token noise.
It cannot perform semantic search.
Even a slight change in chunk splitting can result in zero matches.

Expected Issues in the Standard Chunk

The standard chunk contains the following tokens all at once.

func
process
delta
velocity
vector2
zero
input
is_action_pressed
move_right
move_left
normalized
animatedsprite2d
play
stop
position
clamp
screen_size

All of these tokens must be present together in a single JSONL row, otherwise the search results may be missing.

PoC Simulation

Assume the reference chunk is passed directly to /api/retrieve.

Input:

chunkText = entire _process(delta) function body

The current method converts this input to plainto_tsquery('simple', chunkText).

Expected query characteristics:

func & process & delta & var & velocity & vector2 & zero
& input & is_action_pressed & move_right & move_left
& length & normalized & speed & animatedsprite2d
& play & stop & position & clamp & screen_size

At this point, a problem occurs if the actual related document is split into two chunks as shown below.

docs chunk A:
  Input.is_action_pressed("move_right")
  Input.is_action_pressed("move_left")
  velocity.normalized() * speed
  $AnimatedSprite2D.play()
  $AnimatedSprite2D.stop()

docs chunk B:
  position += velocity * delta
  position = position.clamp(Vector2.ZERO, screen_size)

If a person looks, both A and B are related.

However, if a full-text query demands too many tokens at once, it can result in something like the following.

docs chunk A:
  animatedsprite2d = present
  is_action_pressed = present
  move_left/move_right = present
  clamp = absent
  screen_size = absent
  position.clamp = absent
  => query does not satisfy the whole condition

docs chunk B:
  clamp = present
  screen_size = present
  position = present
  animatedsprite2d = absent or weak
  move_left/move_right = absent
  is_action_pressed = absent
  => query does not satisfy the whole condition

Visible forms of failure:

Search results 0 items  
or return only a very narrow chunk

In this method, it is difficult to distinguish whether a search failure is due to “no basis in the document” or “the query conditions being too strict”.

Logs to Verify in PoC

When testing this alternative, you should not look only at the results.

You need to show the following values together.

1. raw chunkText  
2. query created with `plainto_tsquery`  
3. list of tokens included in the query  
4. number of rows returned  
5. if zero results, which token made it too narrow

Expected Observation:

The longer the raw chunk, the longer the query becomes, and if the related documents are split across multiple chunks, the results are missing.

Judgment

PoC baseline is usable  
Final search strategy is a discard candidate

The judgment of the original note is close to “It’s better to redo it.”