RAG & knowledge base · PostgreSQL and pgvector
CoderConda builds a RAG knowledge base that makes your contracts, minutes and tickets searchable: every answer names its source, and where nothing is written, the system says so. A deadline from an old contract otherwise costs twenty minutes of searching.
An example with an invented corpus. This is what a sourced answer looks like.
What notice period applies in the Meier framework contract?
Answer
Three months to the end of a quarter.
01 — The evidence
The corpus of an invented company: framework contracts, minutes, a quality manual, three years of tickets. For one question the sources contradict each other, for another the answer is nowhere. The system says both instead of smoothing them over.
Three months to the end of a quarter, under § 11 (2) of the 2024 framework contract. For partial services under annex 3, the 2025 addendum names six weeks. Both versions are signed, so the question needs settling before anyone serves notice.
Sources
Annually by a qualified person, plus an extended inspection every four years. That is what the quality manual says, chapter 6.2, as of March 2026. The platform in hall 2 was last inspected on 11 April 2026.
Sources
There is no binding rule on this in the corpus. Three quotes from 2025 with 14 days and two with 30 days turned up, none referring to a policy. That is practice, not a rule. The question goes into the gap report.
Sources
Yes, three days must be held back for the company shutdown, set out in the 2023 works agreement, clause 5. A second source sits in a folder the person asking has no permission for. It was neither read nor quoted.
Sources
02 — The split
The same contract, split two ways. For answer quality this step decides more than the choice of model.
A · Split after 500 characters
Piece 1 · found…renews for twelve months at a time unless terminated with three months’ notice to the end of a quar
Piece 2 · not foundter. By way of exception, for partial services under annex 3…
A period without its condition
The period sits in the first piece, what it refers to in the second. Anyone who finds only the first reads a period without its condition, and gets an answer that is accurate and still wrong.
B · Split along the structure
§ 11 (2) · Term and termination
The contract renews for twelve months at a time unless terminated with three months’ notice to the end of a quarter.
A period with its condition and its source
A passage that stands on its own, with heading, source, version and date. It can be quoted without anyone having to put the original next to it.
03 — The build
For vector search a second database often gets set up, with its own backup and its own permission model. At the volumes of a mid-sized company pgvector in PostgreSQL carries it, and your business data already sits there.
PDF, Word, Confluence, tickets, mail archives, databases. Changed documents are read in again.
PDF · wiki · tickets · SQLPassages with heading, source, date, version and owner. This is where answer quality is decided.
Chunking with metadataPostgreSQL with pgvector. Vectors, full text and your business data in one database, with one backup and one permission model.
PostgreSQL · pgvectorVector and full text combined, then reranked. Names and numbers do not get lost, and neither does meaning.
Hybrid + rerankingEvery row carries its visibility. Anyone who may not open a document does not get it as a source either.
Row level securityWhich questions come in, which stay open, where documents are missing. The report shows what nobody wrote.
Gaps become visibleAt very large corpora and very high load a specialised vector database is faster. Where that line falls for you, I calculate rather than assert.
04 — The offer
What the build costs depends on volume, format and the number of sources. That can only be said after a look at the corpus. A figure before that would be a guess, which is why there is none here.
I read your corpus in, count what is there, and tell you what carries and what needs preparation.
Ingest, split, search and permissions. Quality is measured on fifty real questions, before the build and after every rebuild.
Schema, ingest runs, permission model, test set and documentation. The corpus is yours, and the database is one any developer can read.
Figure
Before the build
I name the amount in writing after the first call, before anything is built.
05 — Objections
Because your business data already sits in a database, and a second system means backup, permissions and operations twice over. pgvector carries the data volumes that mid-sized companies deal with and allows vector search and full text in one query. For very large corpora a specialised database is the better choice. Then I say so and name the threshold in numbers.
For names, numbers and clause references it is precise, which is why it stays part of the system. It fails where the question is worded differently than the document.
| Criterion | Full-text search | RAG system |
|---|---|---|
| Question | Keywords. | Full sentences. |
| Result | A list of documents. | An answer with its sources. |
| Synonyms | Does not find them. | Finds the meaning. |
| Names, numbers | Strong. | Equally, full text stays in the mix. |
| Contradiction | Nobody notices. | Gets named. |
| Gap | Zero results. | “Not written anywhere”, plus a note. |
The model answers exclusively from the retrieved passages and names the source for every statement. If the search finds nothing suitable, the answer is that it is not written anywhere. That is measured against a fixed question set from your everyday work, not against how the demo felt.
Permissions hang on the data. Every passage carries its visibility, and the search runs in the context of the signed-in user. What somebody may not open in the file store appears in no answer either.
Ingest is a repeatable run. Changed files are re-split and replaced, versions stay traceable. A new version replaces the old one, and where two signed versions contradict each other, the answer names both and flags the contradiction.
By a measurement: fifty real questions, checked for whether the right source was among the first hits and whether the answer is correct. The number stands before the build and after every rebuild. Otherwise quality is an opinion.
Spoken in English · subtitles in German, English and Serbian
06 — Who builds it
I am Damjan Savić and built systems for large companies in the corporate world. Today I build the same technology for companies without an IT department of their own. I survey the corpus myself, I decide where it is split, and I check the fifty questions myself.
07 — Appointment
Name me five questions that keep coming up in your company. From those I can see whether your corpus carries and where documents are missing that nobody ever wrote. Pick a time straight from the calendar, no obligation and no sales pressure.
The calendar is hosted by cal.com.
Loading it opens a connection from your browser to cal.com (USA), which transmits your IP address; cal.com may set cookies when it loads. You can also go straight there instead.
