feat: add postgres ingest

[MasterOdin]

PR adds a ingest/postgres_docs.py script that:

1. Clones the https://github.com/postgres/postgres repo locally if it doesn't exist
2. For each version of (14, 15, 16, 17):
   1. Runs ./configure for the version
   2. Runs make html to build HTML pages of the site
   3. Converts each generate HTML file to markdown using markdownify (saved to ingest/build/md)
   4. Create docs.postgres_pages_tmp and docs.postgres_chunks_tmp tables
   5. For each markdown file:
      1. Insert an entry into docs.postgres_pages_tmp
      2. chunk it based on header, and if the chunk is too large splitting it into 7000 token chunks
      3. insert each chunk into `docs.postgres_chunks_tmp
   6. Rename docs.postgres_pages_tmp to docs.postgres_pages and docs.postgres_chunks_tmp to docs.postgres_chunks

The structure of the two tables:

CREATE TABLE docs.postgres_pages (
  id int4 PRIMARY KEY generated by default as identity
  , version int2 NOT NULL
  , url TEXT UNIQUE NOT NULL
  , domain TEXT NOT NULL
  , filename TEXT NOT NULL
  , content_length INTEGER
  , scraped_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
  , chunking_method TEXT DEFAULT 'header'
  , chunks_count INTEGER DEFAULT 0
);

CREATE TABLE IF NOT EXISTS docs.postgres_chunks (
  id int4 PRIMARY KEY generated by default as identity
  , page_id INTEGER REFERENCES docs.postgres_pages(id) ON DELETE CASCADE
  , chunk_index INTEGER NOT NULL
  , sub_chunk_index INTEGER NOT NULL DEFAULT 0
  , content TEXT NOT NULL
  , metadata JSONB
  , embedding vector(1536)
  , created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

Creation of the base tables will be handled by #3 once that's ready to be merged, however for now focusing on just getting the ingest pipelines done, and then will return to that.

I had originally started with the approach taken in https://github.com/timescale/pg-rag, however it had the following issues:

1. It wasn't actually building the docs for 14 and 15 (they lacked target for postgres-full.xml)
2. The postgres-full.xml -> postgres-full.md conversion left with bad headers for the ref entries causing them to be chunked poorly.
3. Getting the proper source_url for each chunk was an annoyingly fraught process (see feat: return sourceUrl for postgres docs #4 for the script I made, which still left about a third of entries without a url), whereas by processing individual HTML pages it was super easy.

---

As part of this PR, I've intentionally omitted pyproject.toml, README.md, etc. where I plan to add those later. If you wish to run this locally, you will need to edit the root .env file to have PG... variables set to connect to the database, and then the dependencies are:

dependencies = [
    "beautifulsoup4>=4.13.5",
    "markdownify>=1.1.0",
    "openai>=1.97.1",
    "psycopg[binary,pool]>=3.2.9",
    "python-dotenv[cli]>=1.1.1",
    "tiktoken>=0.11.0",
]

[jgpruitt]

Can we have a requirements.txt or a pyproject.toml here to make it easy to pull dependencies and run this, please?

[jgpruitt]

This is intended to be run on the host and not in a docker container, right? Can we have a readme with prereqs? I used to run this process in docker b/c I didn't want all that stuff installed on my mac.

[MasterOdin]

Could easily be either. I plan to add a README in a follow-up PR that covers both this and #7. I can also look to add a Dockerfile to run these as well.