Pass 3 of 3 · Deep dive of PHAROSPass 1 · OverviewPass 2 · How it works
Code Observatory / History
What changed, commit by commit.
Selected decisions from the Git history. Dates and sizes come from the commits. A reason appears only where the commit itself records one.
GIT HISTORY Every entry links its commit at the repository. This is a selection, not the whole log.
PHAROS at 975a38dMIT license
15 milestones
storage
Similarity moves into pgvector
The search path stopped pulling every chunk row into Python for cosine similarity and queried an HNSW cosine index instead. The commit records that the old path hung for minutes on a 512 MB host once a real repository was ingested.
60c134c1Replace O(N) Python cosine similarity with pgvector in parent_child_search 1 files, +54 −67service
The knowledge-management product is removed
Annotations, authority, scholarly, quality, collections and auth modules were deleted, along with unused Celery tasks. 25,792 lines left the tree, and the commit states that PHAROS is now a focused code-intelligence backend.
f9858586refactor: Phase 2 — amputate Neo Alexandria 91 files, +147 −25,792service
Module boundaries become a CI gate
Shared utilities moved to app/shared, cross-module imports were cut, and the commit message reports zero isolation violations with the CI gate made hard.
d566d715refactor: Phase 3 — module isolation + Resource column surgery 13 files, +167 −75interface
A real MCP server
PHAROS mounted a native streamable-HTTP MCP server at /mcp, exposing retrieve_context and four graph and search tools at this commit. Existing REST endpoints were left as they were.
572e754cfeat(mcp): speak real MCP — streamable-HTTP server (plan 001) 8 files, +594 −9indexing
Today's core: chunk-level indexing and hybrid retrieval
A new app/codeintel replaced the old repository path. The commit records why: an audit found the old pipeline embedded every symbol but stored only the first symbol's vector, returned the first function of each file, and duplicated rows on re-ingest. Blob-SHA source verification and the conservative call graph first appear here.
fced62c0codeintel: chunk-level indexing, hybrid retrieval and context assembly 44 files, +6,054 −0storage
Migrations squashed, code index schema added
The migration chain was squashed into a baseline, with a second revision adding the tables and indexes the new core needs: repositories state, chunk columns, code_edges and ingestion_jobs, plus HNSW and GIN indexes on PostgreSQL.
3620ffbedb: explicit drivers, honour DATABASE_URL, squash migrations 64 files, +1,331 −279interface
Two tokens, and working MCP on the new core
Authentication became one middleware with admin and read bearer tokens, compared in constant time and failing closed. The read token is limited to GETs and listed read-only POSTs.
f8d37f02api: scoped M2M auth, /api/v1 code-intelligence endpoints, working MCP 12 files, +1,776 −1,689evaluation
Retrieval gets a benchmark
The first retrieval benchmark report was added, with its modes, latency and storage measurements.
7c17c196bench: retrieval quality, latency, throughput and storage measurements 14 files, +13,017 −22service
Free deployment with CPU embeddings
A blueprint was added for a free in-process worker with CPU ONNX embeddings.
0b017297Merge pull request #2 from ram-tewari/deploy/new-prod-onnx 10 files, +255 −14retrieval
Granite embeddings and dense by default
Phase 1 changed the embedding model, the text that gets embedded, the default mode and the parsers. Choices were made on development sets, then measured once on two locked test sets.
a9a80a38Merge pull request #3 from ram-tewari/retrieval-phase1 55 files, +62,932 −138service
Query embedding can run off the host
The edge-onnx backend was added: the API asks a separate machine to embed queries and falls back to its own model when that machine is unreachable.
838154eeMerge pull request #5 from ram-tewari/edge-query-embedding 10 files, +384 −4evaluation
The memory evaluation is locked first
The protocol and its task sets were committed before any memory code existed, fixing the primary measure in advance.
91059434Merge pull request #8 from ram-tewari/phase2-eval-protocol 6 files, +6,248 −1evaluation
Memory fails its locked evaluation
The run failed its primary source-file recall gate and its precision gate at 55% audited. Repository Q&A improved as a secondary result.
bbe807e6Merge pull request #13 from ram-tewari/phase2-locked-run 9 files, +22,891 −4evaluation
Memory claims must prove themselves
Counted claims gained tallies that the grounding check re-verifies, in response to the 55% result. Its precision on new data remains unmeasured.
43d5bd3dMerge pull request #14 from ram-tewari/phase2-evidence 10 files, +774 −101evaluation
The second evaluation stops with no verdict
A second run was stopped at 66 of 80 tasks. The report gives partial numbers and no verdict. This is the commit every source link on this site is pinned to.
975a38d5Merge pull request #18 from ram-tewari/phase2-v2-partial 8 files, +38,983 −7
