Project

Apollo Mission Logs

p_apollo

Documents

142

Chunks

8,214

Vectors

8,214

Storage

29.9 MB

Drop files — or paste a folder. PDF, MD, TXT, DOCX accepted. Bulk OK.

Documents

file

stage

chunks

tokens

size

received

AS-204_review_board_summary.pdf

indexed

84

21,840

2.0 MB

14:02:17

lunar_module_mission_report.pdf

indexed

312

81,120

8.5 MB

14:02:09

flight_director_loop_transcript.txt

embedding

47

12,220

1.2 MB

14:02:01

apollo_13_post_mission_anomaly.pdf

chunking

—

—

5.4 MB

14:01:58

press_kit_apollo_11.pdf

extracting

—

—

1.9 MB

14:01:54

capcom_audio_log_2024_07.txt

failed encoding

—

—

400.4 KB

14:01:31

training_manual_v3.pdf

indexed

156

40,560

3.0 MB

13:58:02

orbital_mechanics_primer.md

indexed

22

5,720

85.9 KB

13:55:44

Chunk inspector AS-204_review_board_summary.pdf — first 5 of 84

#000 261t

A REVIEW OF THE FAILURE OF APOLLO SPACECRAFT COMMAND MODULE 012 ON 27 JANUARY 1967 AT 23:31:04 GMT.

#001 248t

The accident occurred during a pre-launch test designated "plugs-out" while the spacecraft was mounted atop the launch vehicle.

#002 256t

Three astronauts — Lt. Col. Virgil I. Grissom, Lt. Col. Edward H. White II, and Lt. Cmdr. Roger B. Chaffee — perished from asphyxiation.

#003 271t

A bare wire arc beneath the lower equipment bay floor in proximity to environmental control system instrumentation panels.

#004 244t

The pure oxygen atmosphere at 16.7 psi maintained inside the cabin acted as accelerant; combustion propagated within seconds.

DOC

arch — pipeline & memory model

01

Pipeline

file → vector → collection

┌──────┐    ┌─────────┐    ┌────────┐    ┌──────────┐    ┌────────────┐
│ FILE │ ─► │ EXTRACT │ ─► │ CHUNK  │ ─► │  EMBED   │ ─► │   STORE    │
│ pdf  │    │  text   │    │ ~512t  │    │ vec[1024]│    │  qdrant    │
│ md   │    │  utf-8  │    │ +64 ov │    │  ollama  │    │ /proj_id   │
│ txt  │    │         │    │        │    │  qwen3   │    │  cosine    │
└──────┘    └─────────┘    └────────┘    └──────────┘    └────────────┘

▌ ndjson · receive · extract · chunk · embed · indexed · done ─────►

Each upload passes through five stages, every one emitting a typed ndjson event back to the client. The activity feed is a direct render of that stream — there is no polling, no secondary state, no out-of-band status. New stages (graph extraction · dedup · summary) hang off the same wire and surface in the same UI without rewiring.

02

Memory model

git, but for meaning

Most LLM context is ephemeral — pasted into a chat window, used once, lost. embed inverts that premise: a versioned, queryable substrate where institutional knowledge accumulates. Every project is an isolated namespace — one Qdrant collection, one embedding model, one fixed vector dimension. Documents are addressable artifacts; chunks carry provenance; payloads survive every round-trip.

Where git tracks lines and diffs, embed tracks meaning and proximity. Where git log shows commits, the activity feed shows ingestion events. The graph layer (planned) will track derivations between facts the way git tracks derivations between commits — branchable, auditable, blame-able. A central place for context, not a scratch pad.

03

Invariants

load-bearing guarantees

SCOPE Every vector op is keyed by the project's Qdrant collection. Cross-project leakage is structurally impossible.
DIM-LOCK Collection vector size is fixed at create time. Swapping the embedding model mints new collections — old ones reject mismatched-dim upserts.
STREAM All progress is one JSON object per line. No SSE, no polling. Trivially testable with curl.
NO-LC No LangChain. ~150 lines of recursive splitter, Ollama wrapper, and Qdrant fetch wrapper — full control of chunk metadata, retries, and progress events.